[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79NGTZQPCVD3AD8HG9SH4CE":3,"skill-same:01M79NGTZQPCVD3AD8HG9SH4CE":255},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":15,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":13,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M79NGTZQPCVD3AD8HG9SH4CE","skill-development",null,"Create, test, evaluate, improve, repair, and consolidate Codex skills. Use when the user asks to \"create a skill\", \"add a skill to plugin\", \"write a new skill\", \"improve a skill\", \"make a skill for X\", \"turn this into a skill\", \"convert a slash command to a skill\", \"apply a skill template\", \"improve skill description\", \"organize skill content\", \"fix a broken skill\", \"repair a skill\", \"consolidate skills\", \"find duplicate skills\", \"bulletproof a skill\", \"verify skill under pressure\", or \"compliance test a skill\". Also covers skill structure, progressive disclosure, skill categories, testing, compliance testing, and slash-command conversion for Codex plugins.","AndreHahm\u002Fandres-cc-marketplace","AndreHahm","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F80748019?v=4","01M79N2VZJT9MAJ7K695C9VXQ4","01M79N2W06PFX8BSCQ592M8K8K",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: skill-development\ndescription: >-\n  Create, test, evaluate, improve, repair, and consolidate Codex skills. Use when the user\n  asks to \"create a skill\", \"add a skill to plugin\", \"write a new skill\", \"improve a skill\",\n  \"make a skill for X\", \"turn this into a skill\", \"convert a slash command to a skill\", \"apply a\n  skill template\", \"improve skill description\", \"organize\n  skill content\", \"fix a broken skill\", \"repair a skill\", \"consolidate skills\", \"find duplicate\n  skills\", \"bulletproof a skill\", \"verify skill under pressure\", or \"compliance test a skill\".\n  Also covers skill structure, progressive disclosure, skill categories, testing, compliance\n  testing, and slash-command conversion for Codex plugins.\nallowed-tools: Read Write Edit Glob Grep Agent Skill Bash(python:*) Bash(mkdir:*)\n---\n\n# Skill Development for Codex Plugins\n\nCreate, test, measure, and iteratively improve skills using category-aware design, gotchas-driven development, and progressive disclosure. Covers the full lifecycle: from understanding intent through design, testing, improvement, and polish. Also covers slash-command conversion, repair, and consolidation.\n\n## When to Use\n\nThis skill covers the full skill lifecycle — creation, testing, improvement, repair, and consolidation. The Quick Start table below maps your goal to the right entry path.\n\n## Quick Start\n\nIdentify the entry path and jump in:\n\n| Path | When to Use | First Step |\n|---|---|---|\n| **Extract** | \"turn this into a skill\" | Extract tools\u002Fsteps from conversation → write minimal SKILL.md → test on original task |\n| **Greenfield** | \"make a skill for X\" | Start with one concrete task → succeed → extract winning approach into skill |\n| **Audit** | \"improve\u002Faudit this skill\" | `cp -r` snapshot → run Quality Checklist → fix violations |\n| **Convert** | \"convert slash command to skill\" | Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fslash-command-conversion.md` |\n| **Repair** | \"fix\u002Frepair a skill\" | `scripts\u002Frepair_skill.py --diagnose \u003Cskill-name>` |\n| **Consolidate** | \"find duplicate\u002Fconsolidate skills\" | `scripts\u002Fscan_skills.py` → `scripts\u002Fanalyze_similarity.py` |\n\n## When NOT to Use\n\n- **Refining an existing skill only** → use `\u002Fskill-refiner-interactive`\n- **Reviewing skill quality** → use the `skill-reviewer` agent\n- **Quick standalone repair (issue already diagnosed)** → run `scripts\u002Frepair_skill.py` directly; skip the full skill workflow\n- **Standalone empirical benchmarking of an already-built skill** (with_skill vs. baseline pass rates, timing\u002Ftoken metrics, iteration-over-iteration comparison) → use `skill-tester`. This skill's own Phase 3 is scoped to validating a skill *during its own creation\u002Faudit workflow*, not a dedicated benchmark pipeline — don't run both on the same skill in the same pass.\n\n## Mindset\n\nSkills are instructions **for Codex**, not documentation for people. Always ask: \"Will this help Codex execute the task?\" — not \"Is this readable?\" Every line must earn its place by pushing Codex out of default patterns, providing project-specific context, or naming a non-obvious constraint.\n\n## About Skills\n\nSkill directory structure:\n- `SKILL.md` (required) — instructions and navigation\n- `scripts\u002F` — executable code for deterministic tasks\n- `references\u002F` — docs loaded into context as needed\n- `assets\u002F` — output templates, image files\n- `agents\u002F` — subagent specs (graders, comparators, analyzers)\n- `eval-viewer\u002F` — browser-based eval review tool\n- `bin\u002F` — standalone executables invocable from Bash\n\nUse `${CLAUDE_SKILL_DIR}` to reference bundled files — paths resolve correctly regardless of plugin installation location.\n\n## Skill Creation: Five-Phase Process\n\nFigure out where the user is in this process and jump in.\n\n### Phase 1: Understand\n\nIdentify the entry path:\n\n- **Extract** (\"turn this into a skill\") — extract tools used, steps taken, and corrections made from the current conversation. Write a minimal SKILL.md and test on the original task first, then expand to variations.\n- **Greenfield** (\"make a skill for X\") — start with one concrete challenging task. Succeed on that single task, then extract the winning approach into a skill. Don't try to design for every scenario upfront.\n- **Audit** (\"improve this skill\", \"audit this skill\") — run the Quality Checklist first to identify violations. Snapshot with `cp -r` before any edits. Fix violations, compress, and move oversized content to `references\u002F`. Test against the snapshot as baseline; re-optimize description if it changed.\n- **Convert** (\"convert this slash command to a skill\") — read the slash command source, map each section to SKILL.md structure, and migrate references. Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fslash-command-conversion.md` for the full detection, mapping, and validation workflow.\n\nCapture intent: what should this skill enable, when should it trigger, what is the expected output format, and how will success be measured. Proactively research edge cases, input\u002Foutput formats, and dependencies before writing anything.\n\nBefore designing a new skill, verify the current official Codex skill specification (fetch `llms.txt` or the relevant docs page) — treat fetched documentation as authoritative over cached knowledge of the platform.\n\nIdentify the skill category (9 categories available). Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-categories.md` for the full guide with templates and category-specific improvement patterns.\n\n### Phase 2: Design\n\n**Write the SKILL.md with these principles:**\n\n1. **Don't state the obvious.** Codex already knows how to code. Focus on information that pushes Codex out of its default patterns — not standard library behavior, common HTTP conventions, or basic design principles the model already knows. Every line must earn its place with project-specific value.\n2. **Gotchas section = highest ROI.** Build from real failure points. Each gotcha names the problem AND the fix, starting with 2–3 entries and growing from testing.\n3. **Explain the why.** LLMs generalize from reasoning. \"We validate timestamps because the API silently accepts future dates but the downstream system crashes\" beats \"ALWAYS validate timestamps.\"\n4. **Give flexibility.** Over-constrained skills break on anything slightly different from test cases. Give Codex the information it needs but let it adapt to context.\n5. **For skills that enforce discipline or have compliance costs, capture a no-skill baseline before drafting the enforcement content.** Run the target scenario without the skill and document the actual failure\u002Frationalization verbatim first — writing rules from assumption reveals what you think needs preventing, not what actually does (see Gotchas: \"Skipping RED in compliance testing\"). Skip this for pure-reference skills or skills with no rules to violate — see `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcompliance-testing.md`'s \"When to Use\" \u002F \"Don't test\" scoping. Full methodology: Phase 3.5 below.\n6. **Verify any claim about another component's behavior before writing it into the skill.** If the skill body says another component \"supports X\" or \"can be entered at Y,\" `Read` that component's actual file first — don't infer the claim from architectural intent or from what a good design *should* support. A cross-component instruction that turns out to be wrong ships as a defect only a downstream reviewer catches, not something this skill's own design process caught.\n7. **Before presenting a drafted skill at the Design gate, grep the body for every entry on its own `allowed-tools` line and confirm each is actually invoked.** An unused tool grant (declared but never called) is a real R6 least-privilege violation, and it's cheap to self-catch — a 30-second grep — versus expensive to catch later: two skills built in one real session each shipped with an unused tool grant (one an unused `Skill()` call only ever mentioned in prose, one an unused `Bash` scope) that were caught only by the mandatory pre-commit rulebook check, not self-caught during drafting. See `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fallowed-tools.md` for the full least-privilege guidance this check sits alongside.\n8. **When two components in the same build\u002Ffunctional group both gain a new `Skill()` call to each other, trace at least one full round-trip invocation scenario before presenting either at the gate.** Ask: \"what happens if A calls B, and B's own trigger condition for calling A back is also true in that same invocation?\" A cycle designed in this way is invisible to a single-component-at-a-time review — in a real session, exactly this pattern shipped as a Critical circular-dependency bug (component A's new step could nest into component B's new step calling A again, risking duplicate\u002Fnested execution) because the mutual relationship was designed and gate-approved without either side tracing the round-trip. This check belongs here for the single-skill case; for the cross-component functional-group case, `plugin-lifecycle-upstream`'s own Design phase carries the primary version of this check, since that's where the cross-component view actually exists.\n9. **After adding a new policy claim to a file over ~100 lines (\"X is deliberately not supported,\" \"this skill never does Y\"), immediately grep the whole file for the topic keyword before considering the edit done.** A policy\u002Fnegative claim is cheap to contradict elsewhere in a large file and expensive to catch later — in a real session, a new \"no auto-merge path\" line was added to a reference file without grepping the rest of that same file, and a pre-existing \"use `--auto` for trusted changes\" bullet 151 lines below directly contradicted it; the contradiction shipped and was only caught by a later, separate reviewer pass.\n\n**Content distribution (80% Rule):** Core procedural content used in 80%+ of activations stays in SKILL.md. Supplementary content (\u003C20%) moves to `references\u002F`. Never move content solely to reduce line count if execution is affected. See `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-workflow.md` for decision rules and preservation gates. Aim to keep SKILL.md itself under 300 lines for navigation clarity — the hard limit enforced by R13 is 500 lines; when a skill directory contains API documentation, schema tables, or background reference material exceeding ~100 lines, extract it to `references\u002F` as a concrete size-based trigger alongside the 80% Rule.\n\n**Skill placement:** Confirm where the skill will live before creating — use `AskUserQuestion` with options project-scoped (`.Codex\u002Fskills\u002F`), user-space (`~\u002F.Codex\u002Fskills\u002F`), or plugin-scoped (`skills\u002F` in plugin). Never default to user-space without asking — it affects all projects.\n\n**Complex skill considerations:** For error handling, tool scoping, validation, and security review patterns: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcomplex-skills-patterns.md`. For self-containment architecture: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fself-containment-principle.md`. For secrets\u002Fcredentials handling: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fsecrets-and-credentials.md`. For tool scoping (principle of least privilege): `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fallowed-tools.md`. Any skill that uses shell-injected dynamic context must be reviewed as executable content; design it to degrade gracefully when `disableSkillShellExecution` is set.\n\n**Reference linking pattern:** Front-load what the file covers before linking it: `Pattern: [core idea]. See \\`references\u002Ffile.md\\` for [edge cases\u002Fadvanced scenarios].` Without context, Codex loads references out of uncertainty. With it, Codex loads them intentionally.\n\n**Structure the folder** by deciding what each level contains:\n- `scripts\u002F` — helper functions, validation scripts, data fetchers. Bundle when all subagents independently write similar code.\n- `references\u002F` — API docs, detailed specifications. Split by variant for multi-framework support.\n- `workflows\u002F` — multi-step procedural flows invoked from SKILL.md (keep each file under 300 lines). **After writing**: verify no `workflows\u002F` file links to a `references\u002F` file as an action step — such chains force runtime context loads callers don't expect.\n- `assets\u002F` — output templates, image files.\n- `bin\u002F` — standalone executables the Bash tool can invoke by name.\n\nFor platform features (frontmatter fields, string substitutions, hook system), read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fplatform-reference.md`.\n\n**Lifecycle and field semantics:** Skill content is injected into the conversation at invocation and remains active until compaction or session end — write instructions as standing guidance for the full task, not one-time setup text. If a skill expects free-form user input, place `\\$ARGUMENTS` at the intended position in the body; Codex does not append it automatically. Use `disable-model-invocation: true` to block autonomous invocation and `user-invocable: false` only to hide a skill from the slash menu — the latter does not restrict programmatic invocation. Never use `${CLAUDE_PLUGIN_ROOT}` or `${CLAUDE_PLUGIN_DATA}` in prose body text outside code blocks or inline code — these expand at runtime and produce unexpected literal strings in documentation contexts. Skills also support `when_to_use`, `arguments`, `disallowed-tools`, `effort`, `context`, `agent`, `background`, `hooks`, `paths`, and `shell` — full field-by-field schema in `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fschemas.md`.\n\nFor design patterns (sequential workflow, multi-MCP coordination, gotchas structure, progressive disclosure, hooks, composability), read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fdesign-patterns.md`.\n\n**Principle of Lack of Surprise:** Skills must not contain malware, exploit code, or content that facilitates unauthorized access or data exfiltration. A skill's contents should not surprise the user in their intent. Decline requests to create misleading skills or skills designed for malicious purposes.\n\n**Writing patterns** — use these structural templates when defining output formats or examples:\n\n*Output format definition:*\n```markdown\n## Report structure\nALWAYS use this exact template:\n# [Title]\n## Executive summary\n## Key findings\n## Recommendations\n```\n\n*Examples pattern:*\n```markdown\n## Commit message format\n**Example 1:**\nInput: Added user authentication with JWT tokens\nOutput: feat(auth): implement JWT-based authentication\n```\n\n### Phase 3: Test\n\n1. Write 2–3 realistic test prompts. Share with the user for approval. Save to `evals\u002Fevals.json` (schema: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fschemas.md`).\n2. Spawn all runs in one turn — one with-skill, one baseline (no skill, or snapshot of previous version). Launching everything at once lets runs finish around the same time.\n3. **Capture timing** when each subagent task completes — save `timing.json` to the run directory immediately: `{\"total_tokens\": ..., \"duration_ms\": ..., \"total_duration_seconds\": ...}`. This data arrives through the task notification only and isn't persisted elsewhere; process each notification as it arrives.\n4. Draft assertions while runs are in progress. Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Feval-writing-guide.md` for how to write good assertions.\n5. Grade each run using `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fgrader.md`. `grading.json` expectations must use fields `text`, `passed`, and `evidence` — the viewer depends on these exact field names. Aggregate: `python -m scripts.aggregate_benchmark \u003Cworkspace>\u002Fiteration-N --skill-name \u003Cname>` — this writes the iteration's `benchmark.json`\u002F`benchmark.md` and appends a summary row to `\u003Cworkspace>\u002Fbenchmark-log.md`, a cumulative running history across all iterations (append-only — never edit past rows by hand). Then do an analyst pass — read `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fanalyzer.md` for patterns (non-discriminating assertions, high-variance evals, time\u002Ftoken tradeoffs).\n6. Launch eval viewer: `python ${CLAUDE_SKILL_DIR}\u002Feval-viewer\u002Fgenerate_review.py \u003Cworkspace>\u002Fiteration-N --skill-name \u003Cname>`. For iteration 2+: add `--previous-workspace`. The viewer has two tabs — **Outputs** (click through test cases, leave feedback; previous iteration output shown collapsed) and **Benchmark** (pass rates, timing, token usage per configuration). Navigation via prev\u002Fnext or arrow keys; \"Submit All Reviews\" saves `feedback.json`.\n7. After a Codex model update, rerun evals. A pass rate drop signals needed adaptation.\n\n→ Once baseline quality is confirmed, run **Phase 3.5** to verify compliance under adversarial conditions.\n\nDo NOT use `\u002Fskill-test` or other testing skills during eval runs — they conflict with this workflow.\n\n### Phase 3.5: Compliance Testing (TDD for Skills)\n\n**What it tests**: Whether the skill enforces compliance under adversarial conditions — not just whether it improves output quality. Run this after Phase 3 quantitative evals confirm baseline quality.\n\nApply RED-GREEN-REFACTOR to skill documentation:\n\n| TDD Phase | What You Do |\n|-----------|-------------|\n| **RED** | Run scenario WITHOUT skill — document exact failures and rationalizations verbatim |\n| **GREEN** | Write skill addressing specific failures — nothing beyond what you observed |\n| **REFACTOR** | Find new rationalizations, add explicit counters; re-test until no new rationalizations appear |\n\nWrite **pressure scenarios** with 3+ simultaneous pressures (time + sunk cost + authority + exhaustion). Single-pressure scenarios agents resist; multiple pressures reveal real behavior. Force an explicit A\u002FB\u002FC choice — never open-ended.\n\nWhen a scenario fails: capture the rationalization verbatim → add an explicit negation in the rules → add a row to a rationalization table → add a red flag entry → update description with the violation symptom → re-test.\n\nA skill is bulletproof when the agent: chooses the correct option under maximum pressure, cites skill sections, acknowledges the temptation, and meta-testing confirms \"the skill was clear, I should follow it.\"\n\nFor the full methodology — pressure types, scenario templates, rationalization table format, meta-testing protocol (three response types), worked example, common mistakes — read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcompliance-testing.md`.\n\n### Phase 4: Improve\n\n1. Read transcripts, not just final outputs.\n2. Generalize from feedback — fix broadly, not fiddly. Try different metaphors or patterns if a stubborn issue persists.\n3. Keep the prompt lean — remove instructions not pulling their weight. If rewriting finds ALWAYS or NEVER in all-caps, treat it as a signal: reframe as reasoning so the model understands why, rather than imposing rigid constraints.\n4. Bundle repeated code — if all subagents independently write similar helper scripts, bundle the script in `scripts\u002F` and reference from SKILL.md.\n5. Add on-demand hooks if the model strays outside intended boundaries.\n6. Always bump to `iteration-\u003CN+1>\u002F` — never rerun into a previous iteration.\n\nFor blind A\u002FB comparison: `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fcomparator.md`. For benchmark pattern analysis: `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fanalyzer.md`.\n\n### Phase 5: Polish\n\n**Description optimization** — the description is a trigger condition, not a summary. Front-load trigger phrases. `description` should run 80–1024 characters, `when_to_use` up to 512, and the two combined cap at 1536 characters (see `references\u002Fsize-limits.md` for the full tiered thresholds). Describe what the skill does and when to trigger it — not the internal workflow steps; a description that summarizes procedural steps reduces activation accuracy because Codex treats it as already-known detail rather than a trigger to read the body.\n\n**Description formula:** `[Action]. Use when [trigger contexts]. [Scope\u002Fconstraints].` Example: \"Create Codex skills. Use when building new skills, improving existing ones, or converting slash commands. Covers full lifecycle: design, test, improve, polish.\" The exact formula is flexible, but the description must activate Codex on the right input without requiring the skill name to be mentioned — begin with the trigger context or primary use case.\n\nStructure the description as **What + When**: What lists 5–8 specific capabilities using precise action verbs (Validate, Inject, Refactor — not vague terms like \"manage\" or \"handle\"). When defines explicit trigger conditions. For technical skills with file-based activation, add a trigger hint suffix: `(triggers: *.ext, keyword)`.\n\nGenerate 20 trigger eval queries (mix of should-trigger and should-not-trigger). Effective queries are specific: include file paths, personal context, backstory — not generic requests. Save as JSON `[{\"query\": \"...\", \"should_trigger\": true\u002Ffalse}]`. Then review with the user: read `${CLAUDE_SKILL_DIR}\u002Fassets\u002Feval_review.html`, substitute `__EVAL_DATA_PLACEHOLDER__`, `__SKILL_NAME_PLACEHOLDER__`, and `__SKILL_DESCRIPTION_PLACEHOLDER__`, write to a temp file, and open it. The user edits queries and exports `eval_set.json`.\n\nRun the optimization loop from `${CLAUDE_SKILL_DIR}`:\n\n```bash\npython -m scripts.run_loop \\\n  --eval-set \u003Cpath-to-eval.json> \\\n  --skill-path \u003Cpath-to-skill> \\\n  --model \u003Cmodel-id-powering-this-session> \\\n  --max-iterations 5 \\\n  --verbose\n```\n\nApply `best_description` from JSON output to the frontmatter.\n\nFor troubleshooting (doesn't trigger, triggers too often, instructions not followed, large context, frontmatter errors), read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftroubleshooting-guide.md`.\n\nPackage with: `python ${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fpackage_skill.py \u003Cpath\u002Fto\u002Fskill-folder>`\n\nIf the `present_files` tool is available, present the resulting `.skill` file directly to the user. Otherwise, direct the user to the file path for download.\n\n## Plugin-Specific Considerations\n\n### Skill Location in Plugins\n\nPlugin skills live in the plugin's `skills\u002F` directory:\n\n```\nmy-plugin\u002F\n├── .Codex-plugin\u002Fplugin.json\n└── skills\u002F\n    └── my-skill\u002F\n        ├── SKILL.md\n        ├── references\u002F\n        └── scripts\u002F\n```\n\nCreate the structure directly — plugin skills do not use `init_skill.py` (for standalone skills outside a plugin, `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Finit_skill.py \u003Cskill-name> --path \u003Cdir>` scaffolds a template directory):\n\n```bash\nmkdir -p plugin-name\u002Fskills\u002Fskill-name\u002F{references,scripts}\ntouch plugin-name\u002Fskills\u002Fskill-name\u002FSKILL.md\n```\n\n### Auto-Discovery\n\nCodex automatically discovers skills by scanning `skills\u002F` for subdirectories containing `SKILL.md`. Metadata (name + description) is always loaded; the SKILL.md body loads when the skill triggers; references and scripts load as needed.\n\n### No Packaging Needed\n\nPlugin skills are distributed as part of the plugin — not as separate ZIP files. Users get skills when they install the plugin.\n\n### Testing in Plugins\n\n```bash\ncc --plugin-dir \u002Fpath\u002Fto\u002Fplugin\n```\n\n## Skill Repair & Maintenance\n\nSkills fail to load for predictable reasons. Common symptoms: \"Unknown skill\" errors, skills visible on disk but not discoverable, YAML parse failures, structural issues.\n\n### YAML Frontmatter Requirements\n\nYAML validation failures are the most common cause of \"Unknown skill\" errors:\n- `name` must be **kebab-case only** — lowercase letters, numbers, hyphens. No underscores, Title Case, or camelCase.\n- `name` must not contain \"anthropic\" or \"Codex\" (reserved words — rejected silently).\n- `name` must not be a YAML boolean keyword (`on`, `off`, `yes`, `no`, `true`, `false`).\n- Use spaces, not tabs; Unix line endings (LF).\n- Unquoted colons in `description` cause parse errors — wrap the value in quotes.\n\n### Automated Repair\n\n```bash\nscripts\u002Frepair_skill.py --diagnose \u003Cskill-name>   # identify issues without changing anything\nscripts\u002Frepair_skill.py \u003Cskill-name>              # fix automatically\nscripts\u002Frepair_skill.py --list-issues             # audit all skills\n```\n\n### Known Platform Quirks\n\n- **SDK vs CLI:** Skills may not auto-discover when using the Codex Agent SDK — this is a known platform issue. Validate with the CLI first.\n- **Cache:** `\u002Fclear` may not flush skill caches. After adding or modifying a skill, exit and restart Codex completely.\n\n### Prevention\n\n1. Validate with `quick_validate.py` after creation or modification.\n2. Keep directory names and registry entries consistent.\n3. Test skill loading immediately after creation — don't batch validate later.\n\n## Skill Consolidation\n\nIdentify and clean up duplicate or obsolete skills in a growing collection (10+ skills).\n\n### Workflow\n\n```bash\n# Step 1: build inventory\npython scripts\u002Fscan_skills.py --output skill_inventory.json\n\n# Step 2: find merge candidates (Jaccard similarity on triggers\u002Fkeywords)\npython scripts\u002Fanalyze_similarity.py --inventory skill_inventory.json --threshold 0.65\n# outputs: skill_consolidation_report.md + skill_analysis.json\n\n# Step 3: review report and execute merges\u002Farchives manually\n```\n\n### Decision Criteria\n\n**Merge when:** ≥80% similarity, overlapping triggers causing ambiguous activation, or complementary workflows in the same domain.\n\n**Archive when:** Never activated (`activation_count = 0`), unused for 6+ months, tech stack changed, or features removed.\n\n**Archive process:** Move skill directory to `skills\u002Farchive\u002F`, create `ARCHIVED.md` with reason and restoration steps, remove from registry, update `related_skills` references in other skills.\n\n## Writing Style\n\nWrite the entire skill in **imperative\u002Finfinitive form** (verb-first), not second person:\n\n✅ `To create a hook, define the event type.`\n❌ `You should create a hook by defining the event type.`\n\nFrontmatter `description` must use **third person**, prefixed with what the skill does:\n\n✅ `Create Codex skills. Use when the user asks to \"create X\", \"configure Y\"...`\n❌ `Use this skill when you want to create X...`\n\nTarget 1,500–2,000 words for the SKILL.md body. Move detailed content to `references\u002F`.\n\n**Writing philosophy:** Explain the why behind instructions — LLMs generalize from reasoning, not rules. \"We check this because the downstream system crashes on future timestamps\" is more powerful than \"ALWAYS validate timestamps.\" Use theory of mind: understand what the user is trying to accomplish and transmit that understanding into the instructions. Avoid rigid ALWAYS\u002FNEVER directives; if one appears, consider whether the reasoning can be stated instead.\n\n## Communicating with the User\n\nAdapt technical vocabulary to the user's apparent familiarity:\n- \"Evaluation\" and \"benchmark\" are generally acceptable\n- \"JSON\" and \"assertion\" require explanation unless the user has shown clear technical fluency\n- Offer brief definitions when in doubt — non-technical users may be first-time CLI users\n\n## Platform Adaptations\n\n**Codex.ai:** Subagents unavailable — run test cases inline one at a time; skip baseline runs. No browser — present results in conversation instead of launching the viewer. Skip quantitative benchmarking and description optimization (requires `Codex -p` CLI). Packaging still works. When updating an existing skill, preserve the original name and copy to a writable location before editing.\n\n**Cowork:** Subagents work (use parallel runs; fall back to serial on timeouts). No display — use `--static \u003Coutput_path>` with `generate_review.py` to produce a standalone HTML file; share the path as a link. **Generate the eval viewer before reviewing outputs yourself** — present to the user first. Feedback downloads as `feedback.json` when the user clicks \"Submit All Reviews\". Description optimization (`run_loop.py`) works via subprocess; run after the skill is finalized.\n\n## Gotchas\n\n- **Don't use other testing skills during Phase 3.** `\u002Fskill-test` or similar skills will conflict with this skill's eval workflow. Run evals using the steps in Phase 3 directly.\n- **Skipping RED in compliance testing.** Writing compliance rules before watching an agent fail without the skill reveals what YOU think needs preventing, not what actually does. Always run no-skill baseline scenarios first.\n- **Snapshot before improving.** Always `cp -r` the skill before making changes in Phase 4. Without a snapshot, there's no meaningful baseline comparison — the \"before\" is gone.\n- **Create the workspace before spawning subagents.** `mkdir -p \u003Cskill-name>-workspace\u002Fiteration-N\u002F\u003Ceval-name>` upfront. If each subagent tries to create the same parent directory, race conditions produce half-populated directories.\n- **Don't reuse iteration numbers.** When improving, always bump to `iteration-\u003CN+1>\u002F`. Rerunning into a previous iteration silently overwrites the baseline needed for comparison.\n- **`benchmark-log.md` is append-only.** It's the one cumulative file spanning all iterations — never rewrite or delete a past entry by hand. `scripts\u002Faggregate_benchmark.py` appends automatically; only ever add to it, matching the same discipline as not reusing iteration numbers.\n- **Kill the eval viewer.** The viewer process stays alive after review. Forgetting `kill $VIEWER_PID` causes port conflicts or zombie processes on subsequent launches.\n- **Don't over-design upfront.** The biggest time sink is spending 30 minutes on a perfect SKILL.md that turns out to need rewriting after the first eval. Write the minimum, test, then improve.\n- **Description bloat.** If the capability summary in the description exceeds 100 words (quoted trigger phrases excluded), some capabilities belong in the body. Description is a trigger condition, not a manual — move excess to SKILL.md body or `references\u002F`.\n- **Vague file triggers.** Avoid broad patterns like `src\u002F**` or `**\u002F*` for file-based trigger conditions — be surgical. Broad patterns cause over-triggering across unrelated files.\n- **Optional: task tracking for multi-task workflows.** For multi-task skill-creation workflows, using TaskCreate\u002FTaskUpdate\u002FTaskList improves reliability and auditability. This is an internal convention, not a platform requirement.\n- **Substitution variables in documentation examples.** `${CLAUDE_SKILL_DIR}`, `${CLAUDE_PLUGIN_ROOT}`, and similar patterns are expanded by the platform at runtime. Writing them in a SKILL.md code example that is meant to be illustrative — not executed — causes silent expansion that corrupts the example for the reader. Use plain text placeholders (e.g., `\u002Fpath\u002Fto\u002Fskill`, `\u003Cskill-dir>`) in documentation context. Only use `${...}` syntax when the code will genuinely be executed with substitution.\n\n## Testing & Validation\n\n**Verify this skill activates on:**\n- \"create a skill for X\" \u002F \"add a skill to plugin\" \u002F \"write a new skill\"\n- \"improve this skill\" \u002F \"audit this skill\" \u002F \"fix a broken skill\"\n- \"bulletproof a skill\" \u002F \"verify skill under pressure\" \u002F \"compliance test a skill\"\n- \"convert this slash command to a skill\" \u002F \"consolidate my skills\"\n\n**Verify it does NOT activate on:**\n- \"help me debug Python\" \u002F \"review my PR\" \u002F \"write unit tests for X\"\n- \"improve this existing skill\" → use `skill-refiner-interactive` instead\n- \"add a hook to my project\" → use `hook-development` instead\n\n**Quality gates:**\n- `quick_validate.py` passes without errors\n- Description trigger rate ≥80% verified via `run_loop.py`\n- `Codex plugin validate .` passes\n- All Markdown links inside the skill directory resolve — broken links are caught during review, not left for the user to discover\n- All files in the skill directory are reachable from `SKILL.md` directly or via referenced files — an unreferenced supporting file is an orphan smell\n- After creating or modifying a skill, test activation in a fresh session using natural trigger words, not the skill name itself\n\n## Before Packaging\n\nRun `quick_validate.py`, then work through the full pre-release checklist at `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fchecklist.md`. Invoke `plugin-rulebook` to verify naming, language, formatting, and tool-scoping compliance. Before finalizing, verify the current official platform specification for the component type — treat current docs as authoritative over cached or rulebook-cached assumptions.\n\n## Reference Files\n\n### Phase 1–2: Understand & Design\n\n| File | Purpose | When to Read |\n|---|---|---|\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fhow-skills-work.md` | Token loading mechanics, activation internals, selection mechanism | Phase 1 (understand) — deep dive |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-categories.md` | 9 categories with templates and improvement patterns | Phase 1 (identify category) and Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fslash-command-conversion.md` | Detection, mapping, conversion logic, validation | Phase 1 — Convert entry path |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fweb-search-research.md` | Web search research for unfamiliar domains | Phase 1 (understand) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-template.md` | Token-optimized skill template (minimal, \u003C100 lines) | Phase 1 (Greenfield — token-budget focus) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fplatform-reference.md` | Frontmatter fields, string substitutions, hook system, platform gotchas | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fdesign-patterns.md` | Gotchas patterns, progressive disclosure, hooks, setup, composability | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftemplates.md` | Full skill creation templates (Basic, Complex, patterns, description examples) | Phase 2 (design) — comprehensive scaffold |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-workflow.md` | 80% Rule, content distribution, preservation gates | Phase 2 (design) — content placement decisions |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fresource-organization.md` | Content placement decision framework | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcontent-guidelines.md` | Description formulas, phrase library, activation examples | Phase 2 (design) and Phase 5 (polish) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fask-user-question-patterns.md` | AskUserQuestion interaction patterns, decision trees, wizard pattern | Phase 2 (design) — skills with user interviews |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftask-management-patterns.md` | TaskCreate\u002FTaskUpdate instruction patterns for multi-step components | Phase 2 (design) — skills\u002Fagents with 3+ step workflows |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fanti-patterns.md` | Token-wasting, activation, structure, content, and tool-scoping anti-patterns | Phase 2 (design) and Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fadvanced-patterns.md` | Production patterns, skill archetypes, specialized skill designs | Phase 2 (design) and Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcomplex-skills-patterns.md` | Error handling, tool scoping, validation scripts, security review | Phase 2 (design) — complex skills |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fallowed-tools.md` | Tool scoping validation, principle of least privilege | Phase 2 (design) — tool scoping |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fself-containment-principle.md` | External dependency architecture and guidance | Phase 2 (design) — external dependencies |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fsecrets-and-credentials.md` | Secrets detection, git safety, env-var patterns, testing | Phase 2 (design) — skills with credentials |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fsize-limits.md` | Strict size limits reference | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftessl-best-practices.md` | Tessl-style Activation\u002FImplementation standards | Phase 2 and Phase 5 |\n\n### Phase 3–5: Test, Improve & Polish\n\n| File | Purpose | When to Read |\n|---|---|---|\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fschemas.md` | JSON schemas for evals, grading, benchmark, comparison | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Feval-writing-guide.md` | How to write good assertions | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Feval-workflow.md` | Step-by-step eval workflow reference | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fbenchmark.md` | Eval-driven scorecard and iteration table | Phase 3–4 (measure quality) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcompliance-testing.md` | TDD compliance testing: pressure scenarios, rationalization tables, meta-testing protocol | Phase 3.5 — after quantitative evals |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftesting.md` | Trigger rate, eval set schema, regression guide | Phase 3–5 |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Frubric.md` | Tessl-aligned Activation\u002FImplementation grading rubric | Phase 5 (polish) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftroubleshooting-guide.md` | Doesn't trigger, triggers too often, instructions not followed, large context, frontmatter errors | Phase 5 (polish) and debugging |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fchecklist.md` | Comprehensive pre-release validation across all dimensions | Phase 5 (polish) — before packaging |\n| `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fgrader.md` | Evaluate assertions against outputs | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fcomparator.md` | Blind A\u002FB comparison between two outputs | Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fanalyzer.md` | Analyze benchmark patterns and comparison results | Phase 3–4 |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Faggregate_benchmark.py` | Aggregate benchmark results across iterations | Phase 3 step 5 (invoked as `python -m scripts.aggregate_benchmark`) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Frun_loop.py` | Run description optimization loop | Phase 5 (polish) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fquick_validate.py` | Quick skill validation check | Phase 5 (before packaging) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fpackage_skill.py` | Package skill into distributable `.skill` file | Phase 5 (packaging) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Finit_skill.py` | Scaffold template directory for standalone (non-plugin) skills | Phase 1 (Greenfield — non-plugin only) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Frepair_skill.py` | Diagnose and repair skill loading failures | Skill Repair section |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fscan_skills.py` | Inventory all skills with metadata and usage statistics | Skill Consolidation step 1 |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fanalyze_similarity.py` | Find merge candidates by trigger\u002Fkeyword overlap | Skill Consolidation step 2 |\n| `plugin-rulebook` skill | Plugin-level rules — invoke before finalizing to check naming, language, formatting, tool-scoping, and external-reference compliance | Phase 5 (before packaging) |\n| `CHANGELOG.md` | Change history for this skill | Reference |\n\n**Internal dependencies** (not directly invoked — referenced by the scripts above): `scripts\u002Frun_eval.py`, `scripts\u002Fgenerate_report.py`, `scripts\u002Fimprove_description.py`, `scripts\u002Futils.py`, `scripts\u002F__init__.py`, `eval-viewer\u002Fviewer.html`. `LICENSE.txt` — Apache License 2.0.\n\n**Official docs:** [`llms.txt`](https:\u002F\u002Fcode.Codex.com\u002Fdocs\u002Fllms.txt) — fetch for platform spec changes (frontmatter, hooks, manifest). Bundled references cover design and eval methodology; official docs are authoritative for platform syntax.\n","# Skill Development for Codex Plugins\n\nCreate, test, measure, and iteratively improve skills using category-aware design, gotchas-driven development, and progressive disclosure. Covers the full lifecycle: from understanding intent through design, testing, improvement, and polish. Also covers slash-command conversion, repair, and consolidation.\n\n## When to Use\n\nThis skill covers the full skill lifecycle — creation, testing, improvement, repair, and consolidation. The Quick Start table below maps your goal to the right entry path.\n\n## Quick Start\n\nIdentify the entry path and jump in:\n\n| Path | When to Use | First Step |\n|---|---|---|\n| **Extract** | \"turn this into a skill\" | Extract tools\u002Fsteps from conversation → write minimal SKILL.md → test on original task |\n| **Greenfield** | \"make a skill for X\" | Start with one concrete task → succeed → extract winning approach into skill |\n| **Audit** | \"improve\u002Faudit this skill\" | `cp -r` snapshot → run Quality Checklist → fix violations |\n| **Convert** | \"convert slash command to skill\" | Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fslash-command-conversion.md` |\n| **Repair** | \"fix\u002Frepair a skill\" | `scripts\u002Frepair_skill.py --diagnose \u003Cskill-name>` |\n| **Consolidate** | \"find duplicate\u002Fconsolidate skills\" | `scripts\u002Fscan_skills.py` → `scripts\u002Fanalyze_similarity.py` |\n\n## When NOT to Use\n\n- **Refining an existing skill only** → use `\u002Fskill-refiner-interactive`\n- **Reviewing skill quality** → use the `skill-reviewer` agent\n- **Quick standalone repair (issue already diagnosed)** → run `scripts\u002Frepair_skill.py` directly; skip the full skill workflow\n- **Standalone empirical benchmarking of an already-built skill** (with_skill vs. baseline pass rates, timing\u002Ftoken metrics, iteration-over-iteration comparison) → use `skill-tester`. This skill's own Phase 3 is scoped to validating a skill *during its own creation\u002Faudit workflow*, not a dedicated benchmark pipeline — don't run both on the same skill in the same pass.\n\n## Mindset\n\nSkills are instructions **for Codex**, not documentation for people. Always ask: \"Will this help Codex execute the task?\" — not \"Is this readable?\" Every line must earn its place by pushing Codex out of default patterns, providing project-specific context, or naming a non-obvious constraint.\n\n## About Skills\n\nSkill directory structure:\n- `SKILL.md` (required) — instructions and navigation\n- `scripts\u002F` — executable code for deterministic tasks\n- `references\u002F` — docs loaded into context as needed\n- `assets\u002F` — output templates, image files\n- `agents\u002F` — subagent specs (graders, comparators, analyzers)\n- `eval-viewer\u002F` — browser-based eval review tool\n- `bin\u002F` — standalone executables invocable from Bash\n\nUse `${CLAUDE_SKILL_DIR}` to reference bundled files — paths resolve correctly regardless of plugin installation location.\n\n## Skill Creation: Five-Phase Process\n\nFigure out where the user is in this process and jump in.\n\n### Phase 1: Understand\n\nIdentify the entry path:\n\n- **Extract** (\"turn this into a skill\") — extract tools used, steps taken, and corrections made from the current conversation. Write a minimal SKILL.md and test on the original task first, then expand to variations.\n- **Greenfield** (\"make a skill for X\") — start with one concrete challenging task. Succeed on that single task, then extract the winning approach into a skill. Don't try to design for every scenario upfront.\n- **Audit** (\"improve this skill\", \"audit this skill\") — run the Quality Checklist first to identify violations. Snapshot with `cp -r` before any edits. Fix violations, compress, and move oversized content to `references\u002F`. Test against the snapshot as baseline; re-optimize description if it changed.\n- **Convert** (\"convert this slash command to a skill\") — read the slash command source, map each section to SKILL.md structure, and migrate references. Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fslash-command-conversion.md` for the full detection, mapping, and validation workflow.\n\nCapture intent: what should this skill enable, when should it trigger, what is the expected output format, and how will success be measured. Proactively research edge cases, input\u002Foutput formats, and dependencies before writing anything.\n\nBefore designing a new skill, verify the current official Codex skill specification (fetch `llms.txt` or the relevant docs page) — treat fetched documentation as authoritative over cached knowledge of the platform.\n\nIdentify the skill category (9 categories available). Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-categories.md` for the full guide with templates and category-specific improvement patterns.\n\n### Phase 2: Design\n\n**Write the SKILL.md with these principles:**\n\n1. **Don't state the obvious.** Codex already knows how to code. Focus on information that pushes Codex out of its default patterns — not standard library behavior, common HTTP conventions, or basic design principles the model already knows. Every line must earn its place with project-specific value.\n2. **Gotchas section = highest ROI.** Build from real failure points. Each gotcha names the problem AND the fix, starting with 2–3 entries and growing from testing.\n3. **Explain the why.** LLMs generalize from reasoning. \"We validate timestamps because the API silently accepts future dates but the downstream system crashes\" beats \"ALWAYS validate timestamps.\"\n4. **Give flexibility.** Over-constrained skills break on anything slightly different from test cases. Give Codex the information it needs but let it adapt to context.\n5. **For skills that enforce discipline or have compliance costs, capture a no-skill baseline before drafting the enforcement content.** Run the target scenario without the skill and document the actual failure\u002Frationalization verbatim first — writing rules from assumption reveals what you think needs preventing, not what actually does (see Gotchas: \"Skipping RED in compliance testing\"). Skip this for pure-reference skills or skills with no rules to violate — see `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcompliance-testing.md`'s \"When to Use\" \u002F \"Don't test\" scoping. Full methodology: Phase 3.5 below.\n6. **Verify any claim about another component's behavior before writing it into the skill.** If the skill body says another component \"supports X\" or \"can be entered at Y,\" `Read` that component's actual file first — don't infer the claim from architectural intent or from what a good design *should* support. A cross-component instruction that turns out to be wrong ships as a defect only a downstream reviewer catches, not something this skill's own design process caught.\n7. **Before presenting a drafted skill at the Design gate, grep the body for every entry on its own `allowed-tools` line and confirm each is actually invoked.** An unused tool grant (declared but never called) is a real R6 least-privilege violation, and it's cheap to self-catch — a 30-second grep — versus expensive to catch later: two skills built in one real session each shipped with an unused tool grant (one an unused `Skill()` call only ever mentioned in prose, one an unused `Bash` scope) that were caught only by the mandatory pre-commit rulebook check, not self-caught during drafting. See `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fallowed-tools.md` for the full least-privilege guidance this check sits alongside.\n8. **When two components in the same build\u002Ffunctional group both gain a new `Skill()` call to each other, trace at least one full round-trip invocation scenario before presenting either at the gate.** Ask: \"what happens if A calls B, and B's own trigger condition for calling A back is also true in that same invocation?\" A cycle designed in this way is invisible to a single-component-at-a-time review — in a real session, exactly this pattern shipped as a Critical circular-dependency bug (component A's new step could nest into component B's new step calling A again, risking duplicate\u002Fnested execution) because the mutual relationship was designed and gate-approved without either side tracing the round-trip. This check belongs here for the single-skill case; for the cross-component functional-group case, `plugin-lifecycle-upstream`'s own Design phase carries the primary version of this check, since that's where the cross-component view actually exists.\n9. **After adding a new policy claim to a file over ~100 lines (\"X is deliberately not supported,\" \"this skill never does Y\"), immediately grep the whole file for the topic keyword before considering the edit done.** A policy\u002Fnegative claim is cheap to contradict elsewhere in a large file and expensive to catch later — in a real session, a new \"no auto-merge path\" line was added to a reference file without grepping the rest of that same file, and a pre-existing \"use `--auto` for trusted changes\" bullet 151 lines below directly contradicted it; the contradiction shipped and was only caught by a later, separate reviewer pass.\n\n**Content distribution (80% Rule):** Core procedural content used in 80%+ of activations stays in SKILL.md. Supplementary content (\u003C20%) moves to `references\u002F`. Never move content solely to reduce line count if execution is affected. See `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-workflow.md` for decision rules and preservation gates. Aim to keep SKILL.md itself under 300 lines for navigation clarity — the hard limit enforced by R13 is 500 lines; when a skill directory contains API documentation, schema tables, or background reference material exceeding ~100 lines, extract it to `references\u002F` as a concrete size-based trigger alongside the 80% Rule.\n\n**Skill placement:** Confirm where the skill will live before creating — use `AskUserQuestion` with options project-scoped (`.Codex\u002Fskills\u002F`), user-space (`~\u002F.Codex\u002Fskills\u002F`), or plugin-scoped (`skills\u002F` in plugin). Never default to user-space without asking — it affects all projects.\n\n**Complex skill considerations:** For error handling, tool scoping, validation, and security review patterns: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcomplex-skills-patterns.md`. For self-containment architecture: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fself-containment-principle.md`. For secrets\u002Fcredentials handling: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fsecrets-and-credentials.md`. For tool scoping (principle of least privilege): `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fallowed-tools.md`. Any skill that uses shell-injected dynamic context must be reviewed as executable content; design it to degrade gracefully when `disableSkillShellExecution` is set.\n\n**Reference linking pattern:** Front-load what the file covers before linking it: `Pattern: [core idea]. See \\`references\u002Ffile.md\\` for [edge cases\u002Fadvanced scenarios].` Without context, Codex loads references out of uncertainty. With it, Codex loads them intentionally.\n\n**Structure the folder** by deciding what each level contains:\n- `scripts\u002F` — helper functions, validation scripts, data fetchers. Bundle when all subagents independently write similar code.\n- `references\u002F` — API docs, detailed specifications. Split by variant for multi-framework support.\n- `workflows\u002F` — multi-step procedural flows invoked from SKILL.md (keep each file under 300 lines). **After writing**: verify no `workflows\u002F` file links to a `references\u002F` file as an action step — such chains force runtime context loads callers don't expect.\n- `assets\u002F` — output templates, image files.\n- `bin\u002F` — standalone executables the Bash tool can invoke by name.\n\nFor platform features (frontmatter fields, string substitutions, hook system), read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fplatform-reference.md`.\n\n**Lifecycle and field semantics:** Skill content is injected into the conversation at invocation and remains active until compaction or session end — write instructions as standing guidance for the full task, not one-time setup text. If a skill expects free-form user input, place `\\$ARGUMENTS` at the intended position in the body; Codex does not append it automatically. Use `disable-model-invocation: true` to block autonomous invocation and `user-invocable: false` only to hide a skill from the slash menu — the latter does not restrict programmatic invocation. Never use `${CLAUDE_PLUGIN_ROOT}` or `${CLAUDE_PLUGIN_DATA}` in prose body text outside code blocks or inline code — these expand at runtime and produce unexpected literal strings in documentation contexts. Skills also support `when_to_use`, `arguments`, `disallowed-tools`, `effort`, `context`, `agent`, `background`, `hooks`, `paths`, and `shell` — full field-by-field schema in `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fschemas.md`.\n\nFor design patterns (sequential workflow, multi-MCP coordination, gotchas structure, progressive disclosure, hooks, composability), read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fdesign-patterns.md`.\n\n**Principle of Lack of Surprise:** Skills must not contain malware, exploit code, or content that facilitates unauthorized access or data exfiltration. A skill's contents should not surprise the user in their intent. Decline requests to create misleading skills or skills designed for malicious purposes.\n\n**Writing patterns** — use these structural templates when defining output formats or examples:\n\n*Output format definition:*\n```markdown\n## Report structure\nALWAYS use this exact template:\n# [Title]\n## Executive summary\n## Key findings\n## Recommendations\n```\n\n*Examples pattern:*\n```markdown\n## Commit message format\n**Example 1:**\nInput: Added user authentication with JWT tokens\nOutput: feat(auth): implement JWT-based authentication\n```\n\n### Phase 3: Test\n\n1. Write 2–3 realistic test prompts. Share with the user for approval. Save to `evals\u002Fevals.json` (schema: `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fschemas.md`).\n2. Spawn all runs in one turn — one with-skill, one baseline (no skill, or snapshot of previous version). Launching everything at once lets runs finish around the same time.\n3. **Capture timing** when each subagent task completes — save `timing.json` to the run directory immediately: `{\"total_tokens\": ..., \"duration_ms\": ..., \"total_duration_seconds\": ...}`. This data arrives through the task notification only and isn't persisted elsewhere; process each notification as it arrives.\n4. Draft assertions while runs are in progress. Read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Feval-writing-guide.md` for how to write good assertions.\n5. Grade each run using `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fgrader.md`. `grading.json` expectations must use fields `text`, `passed`, and `evidence` — the viewer depends on these exact field names. Aggregate: `python -m scripts.aggregate_benchmark \u003Cworkspace>\u002Fiteration-N --skill-name \u003Cname>` — this writes the iteration's `benchmark.json`\u002F`benchmark.md` and appends a summary row to `\u003Cworkspace>\u002Fbenchmark-log.md`, a cumulative running history across all iterations (append-only — never edit past rows by hand). Then do an analyst pass — read `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fanalyzer.md` for patterns (non-discriminating assertions, high-variance evals, time\u002Ftoken tradeoffs).\n6. Launch eval viewer: `python ${CLAUDE_SKILL_DIR}\u002Feval-viewer\u002Fgenerate_review.py \u003Cworkspace>\u002Fiteration-N --skill-name \u003Cname>`. For iteration 2+: add `--previous-workspace`. The viewer has two tabs — **Outputs** (click through test cases, leave feedback; previous iteration output shown collapsed) and **Benchmark** (pass rates, timing, token usage per configuration). Navigation via prev\u002Fnext or arrow keys; \"Submit All Reviews\" saves `feedback.json`.\n7. After a Codex model update, rerun evals. A pass rate drop signals needed adaptation.\n\n→ Once baseline quality is confirmed, run **Phase 3.5** to verify compliance under adversarial conditions.\n\nDo NOT use `\u002Fskill-test` or other testing skills during eval runs — they conflict with this workflow.\n\n### Phase 3.5: Compliance Testing (TDD for Skills)\n\n**What it tests**: Whether the skill enforces compliance under adversarial conditions — not just whether it improves output quality. Run this after Phase 3 quantitative evals confirm baseline quality.\n\nApply RED-GREEN-REFACTOR to skill documentation:\n\n| TDD Phase | What You Do |\n|-----------|-------------|\n| **RED** | Run scenario WITHOUT skill — document exact failures and rationalizations verbatim |\n| **GREEN** | Write skill addressing specific failures — nothing beyond what you observed |\n| **REFACTOR** | Find new rationalizations, add explicit counters; re-test until no new rationalizations appear |\n\nWrite **pressure scenarios** with 3+ simultaneous pressures (time + sunk cost + authority + exhaustion). Single-pressure scenarios agents resist; multiple pressures reveal real behavior. Force an explicit A\u002FB\u002FC choice — never open-ended.\n\nWhen a scenario fails: capture the rationalization verbatim → add an explicit negation in the rules → add a row to a rationalization table → add a red flag entry → update description with the violation symptom → re-test.\n\nA skill is bulletproof when the agent: chooses the correct option under maximum pressure, cites skill sections, acknowledges the temptation, and meta-testing confirms \"the skill was clear, I should follow it.\"\n\nFor the full methodology — pressure types, scenario templates, rationalization table format, meta-testing protocol (three response types), worked example, common mistakes — read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcompliance-testing.md`.\n\n### Phase 4: Improve\n\n1. Read transcripts, not just final outputs.\n2. Generalize from feedback — fix broadly, not fiddly. Try different metaphors or patterns if a stubborn issue persists.\n3. Keep the prompt lean — remove instructions not pulling their weight. If rewriting finds ALWAYS or NEVER in all-caps, treat it as a signal: reframe as reasoning so the model understands why, rather than imposing rigid constraints.\n4. Bundle repeated code — if all subagents independently write similar helper scripts, bundle the script in `scripts\u002F` and reference from SKILL.md.\n5. Add on-demand hooks if the model strays outside intended boundaries.\n6. Always bump to `iteration-\u003CN+1>\u002F` — never rerun into a previous iteration.\n\nFor blind A\u002FB comparison: `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fcomparator.md`. For benchmark pattern analysis: `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fanalyzer.md`.\n\n### Phase 5: Polish\n\n**Description optimization** — the description is a trigger condition, not a summary. Front-load trigger phrases. `description` should run 80–1024 characters, `when_to_use` up to 512, and the two combined cap at 1536 characters (see `references\u002Fsize-limits.md` for the full tiered thresholds). Describe what the skill does and when to trigger it — not the internal workflow steps; a description that summarizes procedural steps reduces activation accuracy because Codex treats it as already-known detail rather than a trigger to read the body.\n\n**Description formula:** `[Action]. Use when [trigger contexts]. [Scope\u002Fconstraints].` Example: \"Create Codex skills. Use when building new skills, improving existing ones, or converting slash commands. Covers full lifecycle: design, test, improve, polish.\" The exact formula is flexible, but the description must activate Codex on the right input without requiring the skill name to be mentioned — begin with the trigger context or primary use case.\n\nStructure the description as **What + When**: What lists 5–8 specific capabilities using precise action verbs (Validate, Inject, Refactor — not vague terms like \"manage\" or \"handle\"). When defines explicit trigger conditions. For technical skills with file-based activation, add a trigger hint suffix: `(triggers: *.ext, keyword)`.\n\nGenerate 20 trigger eval queries (mix of should-trigger and should-not-trigger). Effective queries are specific: include file paths, personal context, backstory — not generic requests. Save as JSON `[{\"query\": \"...\", \"should_trigger\": true\u002Ffalse}]`. Then review with the user: read `${CLAUDE_SKILL_DIR}\u002Fassets\u002Feval_review.html`, substitute `__EVAL_DATA_PLACEHOLDER__`, `__SKILL_NAME_PLACEHOLDER__`, and `__SKILL_DESCRIPTION_PLACEHOLDER__`, write to a temp file, and open it. The user edits queries and exports `eval_set.json`.\n\nRun the optimization loop from `${CLAUDE_SKILL_DIR}`:\n\n```bash\npython -m scripts.run_loop \\\n  --eval-set \u003Cpath-to-eval.json> \\\n  --skill-path \u003Cpath-to-skill> \\\n  --model \u003Cmodel-id-powering-this-session> \\\n  --max-iterations 5 \\\n  --verbose\n```\n\nApply `best_description` from JSON output to the frontmatter.\n\nFor troubleshooting (doesn't trigger, triggers too often, instructions not followed, large context, frontmatter errors), read `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftroubleshooting-guide.md`.\n\nPackage with: `python ${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fpackage_skill.py \u003Cpath\u002Fto\u002Fskill-folder>`\n\nIf the `present_files` tool is available, present the resulting `.skill` file directly to the user. Otherwise, direct the user to the file path for download.\n\n## Plugin-Specific Considerations\n\n### Skill Location in Plugins\n\nPlugin skills live in the plugin's `skills\u002F` directory:\n\n```\nmy-plugin\u002F\n├── .Codex-plugin\u002Fplugin.json\n└── skills\u002F\n    └── my-skill\u002F\n        ├── SKILL.md\n        ├── references\u002F\n        └── scripts\u002F\n```\n\nCreate the structure directly — plugin skills do not use `init_skill.py` (for standalone skills outside a plugin, `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Finit_skill.py \u003Cskill-name> --path \u003Cdir>` scaffolds a template directory):\n\n```bash\nmkdir -p plugin-name\u002Fskills\u002Fskill-name\u002F{references,scripts}\ntouch plugin-name\u002Fskills\u002Fskill-name\u002FSKILL.md\n```\n\n### Auto-Discovery\n\nCodex automatically discovers skills by scanning `skills\u002F` for subdirectories containing `SKILL.md`. Metadata (name + description) is always loaded; the SKILL.md body loads when the skill triggers; references and scripts load as needed.\n\n### No Packaging Needed\n\nPlugin skills are distributed as part of the plugin — not as separate ZIP files. Users get skills when they install the plugin.\n\n### Testing in Plugins\n\n```bash\ncc --plugin-dir \u002Fpath\u002Fto\u002Fplugin\n```\n\n## Skill Repair & Maintenance\n\nSkills fail to load for predictable reasons. Common symptoms: \"Unknown skill\" errors, skills visible on disk but not discoverable, YAML parse failures, structural issues.\n\n### YAML Frontmatter Requirements\n\nYAML validation failures are the most common cause of \"Unknown skill\" errors:\n- `name` must be **kebab-case only** — lowercase letters, numbers, hyphens. No underscores, Title Case, or camelCase.\n- `name` must not contain \"anthropic\" or \"Codex\" (reserved words — rejected silently).\n- `name` must not be a YAML boolean keyword (`on`, `off`, `yes`, `no`, `true`, `false`).\n- Use spaces, not tabs; Unix line endings (LF).\n- Unquoted colons in `description` cause parse errors — wrap the value in quotes.\n\n### Automated Repair\n\n```bash\nscripts\u002Frepair_skill.py --diagnose \u003Cskill-name>   # identify issues without changing anything\nscripts\u002Frepair_skill.py \u003Cskill-name>              # fix automatically\nscripts\u002Frepair_skill.py --list-issues             # audit all skills\n```\n\n### Known Platform Quirks\n\n- **SDK vs CLI:** Skills may not auto-discover when using the Codex Agent SDK — this is a known platform issue. Validate with the CLI first.\n- **Cache:** `\u002Fclear` may not flush skill caches. After adding or modifying a skill, exit and restart Codex completely.\n\n### Prevention\n\n1. Validate with `quick_validate.py` after creation or modification.\n2. Keep directory names and registry entries consistent.\n3. Test skill loading immediately after creation — don't batch validate later.\n\n## Skill Consolidation\n\nIdentify and clean up duplicate or obsolete skills in a growing collection (10+ skills).\n\n### Workflow\n\n```bash\n# Step 1: build inventory\npython scripts\u002Fscan_skills.py --output skill_inventory.json\n\n# Step 2: find merge candidates (Jaccard similarity on triggers\u002Fkeywords)\npython scripts\u002Fanalyze_similarity.py --inventory skill_inventory.json --threshold 0.65\n# outputs: skill_consolidation_report.md + skill_analysis.json\n\n# Step 3: review report and execute merges\u002Farchives manually\n```\n\n### Decision Criteria\n\n**Merge when:** ≥80% similarity, overlapping triggers causing ambiguous activation, or complementary workflows in the same domain.\n\n**Archive when:** Never activated (`activation_count = 0`), unused for 6+ months, tech stack changed, or features removed.\n\n**Archive process:** Move skill directory to `skills\u002Farchive\u002F`, create `ARCHIVED.md` with reason and restoration steps, remove from registry, update `related_skills` references in other skills.\n\n## Writing Style\n\nWrite the entire skill in **imperative\u002Finfinitive form** (verb-first), not second person:\n\n✅ `To create a hook, define the event type.`\n❌ `You should create a hook by defining the event type.`\n\nFrontmatter `description` must use **third person**, prefixed with what the skill does:\n\n✅ `Create Codex skills. Use when the user asks to \"create X\", \"configure Y\"...`\n❌ `Use this skill when you want to create X...`\n\nTarget 1,500–2,000 words for the SKILL.md body. Move detailed content to `references\u002F`.\n\n**Writing philosophy:** Explain the why behind instructions — LLMs generalize from reasoning, not rules. \"We check this because the downstream system crashes on future timestamps\" is more powerful than \"ALWAYS validate timestamps.\" Use theory of mind: understand what the user is trying to accomplish and transmit that understanding into the instructions. Avoid rigid ALWAYS\u002FNEVER directives; if one appears, consider whether the reasoning can be stated instead.\n\n## Communicating with the User\n\nAdapt technical vocabulary to the user's apparent familiarity:\n- \"Evaluation\" and \"benchmark\" are generally acceptable\n- \"JSON\" and \"assertion\" require explanation unless the user has shown clear technical fluency\n- Offer brief definitions when in doubt — non-technical users may be first-time CLI users\n\n## Platform Adaptations\n\n**Codex.ai:** Subagents unavailable — run test cases inline one at a time; skip baseline runs. No browser — present results in conversation instead of launching the viewer. Skip quantitative benchmarking and description optimization (requires `Codex -p` CLI). Packaging still works. When updating an existing skill, preserve the original name and copy to a writable location before editing.\n\n**Cowork:** Subagents work (use parallel runs; fall back to serial on timeouts). No display — use `--static \u003Coutput_path>` with `generate_review.py` to produce a standalone HTML file; share the path as a link. **Generate the eval viewer before reviewing outputs yourself** — present to the user first. Feedback downloads as `feedback.json` when the user clicks \"Submit All Reviews\". Description optimization (`run_loop.py`) works via subprocess; run after the skill is finalized.\n\n## Gotchas\n\n- **Don't use other testing skills during Phase 3.** `\u002Fskill-test` or similar skills will conflict with this skill's eval workflow. Run evals using the steps in Phase 3 directly.\n- **Skipping RED in compliance testing.** Writing compliance rules before watching an agent fail without the skill reveals what YOU think needs preventing, not what actually does. Always run no-skill baseline scenarios first.\n- **Snapshot before improving.** Always `cp -r` the skill before making changes in Phase 4. Without a snapshot, there's no meaningful baseline comparison — the \"before\" is gone.\n- **Create the workspace before spawning subagents.** `mkdir -p \u003Cskill-name>-workspace\u002Fiteration-N\u002F\u003Ceval-name>` upfront. If each subagent tries to create the same parent directory, race conditions produce half-populated directories.\n- **Don't reuse iteration numbers.** When improving, always bump to `iteration-\u003CN+1>\u002F`. Rerunning into a previous iteration silently overwrites the baseline needed for comparison.\n- **`benchmark-log.md` is append-only.** It's the one cumulative file spanning all iterations — never rewrite or delete a past entry by hand. `scripts\u002Faggregate_benchmark.py` appends automatically; only ever add to it, matching the same discipline as not reusing iteration numbers.\n- **Kill the eval viewer.** The viewer process stays alive after review. Forgetting `kill $VIEWER_PID` causes port conflicts or zombie processes on subsequent launches.\n- **Don't over-design upfront.** The biggest time sink is spending 30 minutes on a perfect SKILL.md that turns out to need rewriting after the first eval. Write the minimum, test, then improve.\n- **Description bloat.** If the capability summary in the description exceeds 100 words (quoted trigger phrases excluded), some capabilities belong in the body. Description is a trigger condition, not a manual — move excess to SKILL.md body or `references\u002F`.\n- **Vague file triggers.** Avoid broad patterns like `src\u002F**` or `**\u002F*` for file-based trigger conditions — be surgical. Broad patterns cause over-triggering across unrelated files.\n- **Optional: task tracking for multi-task workflows.** For multi-task skill-creation workflows, using TaskCreate\u002FTaskUpdate\u002FTaskList improves reliability and auditability. This is an internal convention, not a platform requirement.\n- **Substitution variables in documentation examples.** `${CLAUDE_SKILL_DIR}`, `${CLAUDE_PLUGIN_ROOT}`, and similar patterns are expanded by the platform at runtime. Writing them in a SKILL.md code example that is meant to be illustrative — not executed — causes silent expansion that corrupts the example for the reader. Use plain text placeholders (e.g., `\u002Fpath\u002Fto\u002Fskill`, `\u003Cskill-dir>`) in documentation context. Only use `${...}` syntax when the code will genuinely be executed with substitution.\n\n## Testing & Validation\n\n**Verify this skill activates on:**\n- \"create a skill for X\" \u002F \"add a skill to plugin\" \u002F \"write a new skill\"\n- \"improve this skill\" \u002F \"audit this skill\" \u002F \"fix a broken skill\"\n- \"bulletproof a skill\" \u002F \"verify skill under pressure\" \u002F \"compliance test a skill\"\n- \"convert this slash command to a skill\" \u002F \"consolidate my skills\"\n\n**Verify it does NOT activate on:**\n- \"help me debug Python\" \u002F \"review my PR\" \u002F \"write unit tests for X\"\n- \"improve this existing skill\" → use `skill-refiner-interactive` instead\n- \"add a hook to my project\" → use `hook-development` instead\n\n**Quality gates:**\n- `quick_validate.py` passes without errors\n- Description trigger rate ≥80% verified via `run_loop.py`\n- `Codex plugin validate .` passes\n- All Markdown links inside the skill directory resolve — broken links are caught during review, not left for the user to discover\n- All files in the skill directory are reachable from `SKILL.md` directly or via referenced files — an unreferenced supporting file is an orphan smell\n- After creating or modifying a skill, test activation in a fresh session using natural trigger words, not the skill name itself\n\n## Before Packaging\n\nRun `quick_validate.py`, then work through the full pre-release checklist at `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fchecklist.md`. Invoke `plugin-rulebook` to verify naming, language, formatting, and tool-scoping compliance. Before finalizing, verify the current official platform specification for the component type — treat current docs as authoritative over cached or rulebook-cached assumptions.\n\n## Reference Files\n\n### Phase 1–2: Understand & Design\n\n| File | Purpose | When to Read |\n|---|---|---|\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fhow-skills-work.md` | Token loading mechanics, activation internals, selection mechanism | Phase 1 (understand) — deep dive |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-categories.md` | 9 categories with templates and improvement patterns | Phase 1 (identify category) and Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fslash-command-conversion.md` | Detection, mapping, conversion logic, validation | Phase 1 — Convert entry path |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fweb-search-research.md` | Web search research for unfamiliar domains | Phase 1 (understand) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-template.md` | Token-optimized skill template (minimal, \u003C100 lines) | Phase 1 (Greenfield — token-budget focus) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fplatform-reference.md` | Frontmatter fields, string substitutions, hook system, platform gotchas | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fdesign-patterns.md` | Gotchas patterns, progressive disclosure, hooks, setup, composability | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftemplates.md` | Full skill creation templates (Basic, Complex, patterns, description examples) | Phase 2 (design) — comprehensive scaffold |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fskill-workflow.md` | 80% Rule, content distribution, preservation gates | Phase 2 (design) — content placement decisions |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fresource-organization.md` | Content placement decision framework | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcontent-guidelines.md` | Description formulas, phrase library, activation examples | Phase 2 (design) and Phase 5 (polish) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fask-user-question-patterns.md` | AskUserQuestion interaction patterns, decision trees, wizard pattern | Phase 2 (design) — skills with user interviews |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftask-management-patterns.md` | TaskCreate\u002FTaskUpdate instruction patterns for multi-step components | Phase 2 (design) — skills\u002Fagents with 3+ step workflows |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fanti-patterns.md` | Token-wasting, activation, structure, content, and tool-scoping anti-patterns | Phase 2 (design) and Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fadvanced-patterns.md` | Production patterns, skill archetypes, specialized skill designs | Phase 2 (design) and Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcomplex-skills-patterns.md` | Error handling, tool scoping, validation scripts, security review | Phase 2 (design) — complex skills |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fallowed-tools.md` | Tool scoping validation, principle of least privilege | Phase 2 (design) — tool scoping |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fself-containment-principle.md` | External dependency architecture and guidance | Phase 2 (design) — external dependencies |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fsecrets-and-credentials.md` | Secrets detection, git safety, env-var patterns, testing | Phase 2 (design) — skills with credentials |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fsize-limits.md` | Strict size limits reference | Phase 2 (design) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftessl-best-practices.md` | Tessl-style Activation\u002FImplementation standards | Phase 2 and Phase 5 |\n\n### Phase 3–5: Test, Improve & Polish\n\n| File | Purpose | When to Read |\n|---|---|---|\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fschemas.md` | JSON schemas for evals, grading, benchmark, comparison | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Feval-writing-guide.md` | How to write good assertions | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Feval-workflow.md` | Step-by-step eval workflow reference | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fbenchmark.md` | Eval-driven scorecard and iteration table | Phase 3–4 (measure quality) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fcompliance-testing.md` | TDD compliance testing: pressure scenarios, rationalization tables, meta-testing protocol | Phase 3.5 — after quantitative evals |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftesting.md` | Trigger rate, eval set schema, regression guide | Phase 3–5 |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Frubric.md` | Tessl-aligned Activation\u002FImplementation grading rubric | Phase 5 (polish) |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Ftroubleshooting-guide.md` | Doesn't trigger, triggers too often, instructions not followed, large context, frontmatter errors | Phase 5 (polish) and debugging |\n| `${CLAUDE_SKILL_DIR}\u002Freferences\u002Fchecklist.md` | Comprehensive pre-release validation across all dimensions | Phase 5 (polish) — before packaging |\n| `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fgrader.md` | Evaluate assertions against outputs | Phase 3 (test) |\n| `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fcomparator.md` | Blind A\u002FB comparison between two outputs | Phase 4 (improve) |\n| `${CLAUDE_SKILL_DIR}\u002Fagents\u002Fanalyzer.md` | Analyze benchmark patterns and comparison results | Phase 3–4 |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Faggregate_benchmark.py` | Aggregate benchmark results across iterations | Phase 3 step 5 (invoked as `python -m scripts.aggregate_benchmark`) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Frun_loop.py` | Run description optimization loop | Phase 5 (polish) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fquick_validate.py` | Quick skill validation check | Phase 5 (before packaging) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fpackage_skill.py` | Package skill into distributable `.skill` file | Phase 5 (packaging) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Finit_skill.py` | Scaffold template directory for standalone (non-plugin) skills | Phase 1 (Greenfield — non-plugin only) |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Frepair_skill.py` | Diagnose and repair skill loading failures | Skill Repair section |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fscan_skills.py` | Inventory all skills with metadata and usage statistics | Skill Consolidation step 1 |\n| `${CLAUDE_SKILL_DIR}\u002Fscripts\u002Fanalyze_similarity.py` | Find merge candidates by trigger\u002Fkeyword overlap | Skill Consolidation step 2 |\n| `plugin-rulebook` skill | Plugin-level rules — invoke before finalizing to check naming, language, formatting, tool-scoping, and external-reference compliance | Phase 5 (before packaging) |\n| `CHANGELOG.md` | Change history for this skill | Reference |\n\n**Internal dependencies** (not directly invoked — referenced by the scripts above): `scripts\u002Frun_eval.py`, `scripts\u002Fgenerate_report.py`, `scripts\u002Fimprove_description.py`, `scripts\u002Futils.py`, `scripts\u002F__init__.py`, `eval-viewer\u002Fviewer.html`. `LICENSE.txt` — Apache License 2.0.\n\n**Official docs:** [`llms.txt`](https:\u002F\u002Fcode.Codex.com\u002Fdocs\u002Fllms.txt) — fetch for platform spec changes (frontmatter, hooks, manifest). Bundled references cover design and eval methodology; official docs are authoritative for platform syntax.\n",{"name":5,"description":7,"allowed-tools":24},"Read Write Edit Glob Grep Agent Skill Bash(python:*) Bash(mkdir:*)","main","1bf12900cac6f28e346abd2912bdb78d306de723",".agents\u002Fskills\u002Fskill-development","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace","https:\u002F\u002Fgithub.com\u002FAndreHahm","2026-11-15 04:26:16",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add AndreHahm\u002Fandres-cc-marketplace --skill skill-development",{"tool":37,"command":38},"bunx","bunx skills add AndreHahm\u002Fandres-cc-marketplace --skill skill-development",{"tool":40,"command":41},"pnpm","pnpm dlx skills add AndreHahm\u002Fandres-cc-marketplace --skill skill-development","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002FSKILL.md",true,[45,51,55,59,65,69,75,79,84,88,92,96,100,104,108,112,116,120,124,128,132,136,140,144,148,152,156,160,164,168,172,176,180,184,188,192,196,200,204,207,211,215,219,223,227,231,235,239,243,247,251],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"agents\u002Fanalyzer.md","REFERENCE","参考文档",3685,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fagents\u002Fanalyzer.md",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"agents\u002Fcomparator.md",2326,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fagents\u002Fcomparator.md",{"path":56,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":57,"sourceHtmlUrl":58},"agents\u002Fgrader.md",3979,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fagents\u002Fgrader.md",{"path":60,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":63,"sourceHtmlUrl":64},"assets\u002Feval_review.html","ASSET","资产文件",7058,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fassets\u002Feval_review.html",{"path":66,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":67,"sourceHtmlUrl":68},"CHANGELOG.md",441,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002FCHANGELOG.md",{"path":70,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":73,"sourceHtmlUrl":74},"eval-viewer\u002Fgenerate_review.py","SCRIPT","脚本",16463,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Feval-viewer\u002Fgenerate_review.py",{"path":76,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":77,"sourceHtmlUrl":78},"eval-viewer\u002Fviewer.html",44998,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Feval-viewer\u002Fviewer.html",{"path":80,"resourceType":81,"resourceTypeDesc":82,"sizeBytes":83,"sourceHtmlUrl":15},"LICENSE.txt","LICENSE","许可证",11358,{"path":85,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":86,"sourceHtmlUrl":87},"references\u002Fadvanced-patterns.md",10211,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fadvanced-patterns.md",{"path":89,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":90,"sourceHtmlUrl":91},"references\u002Fallowed-tools.md",3426,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fallowed-tools.md",{"path":93,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":94,"sourceHtmlUrl":95},"references\u002Fanti-patterns.md",10790,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fanti-patterns.md",{"path":97,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":98,"sourceHtmlUrl":99},"references\u002Fask-user-question-patterns.md",17243,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fask-user-question-patterns.md",{"path":101,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":102,"sourceHtmlUrl":103},"references\u002Fbenchmark.md",2101,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fbenchmark.md",{"path":105,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":106,"sourceHtmlUrl":107},"references\u002Fchecklist.md",14116,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fchecklist.md",{"path":109,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":110,"sourceHtmlUrl":111},"references\u002Fcomplex-skills-patterns.md",3679,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fcomplex-skills-patterns.md",{"path":113,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":114,"sourceHtmlUrl":115},"references\u002Fcompliance-testing.md",11744,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fcompliance-testing.md",{"path":117,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":118,"sourceHtmlUrl":119},"references\u002Fcontent-guidelines.md",6021,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fcontent-guidelines.md",{"path":121,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":122,"sourceHtmlUrl":123},"references\u002Fdesign-patterns.md",21748,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fdesign-patterns.md",{"path":125,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":126,"sourceHtmlUrl":127},"references\u002Feval-workflow.md",4244,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Feval-workflow.md",{"path":129,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":130,"sourceHtmlUrl":131},"references\u002Feval-writing-guide.md",4391,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Feval-writing-guide.md",{"path":133,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":134,"sourceHtmlUrl":135},"references\u002Fhow-skills-work.md",4126,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fhow-skills-work.md",{"path":137,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":138,"sourceHtmlUrl":139},"references\u002Fplatform-reference.md",11140,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fplatform-reference.md",{"path":141,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":142,"sourceHtmlUrl":143},"references\u002Fresource-organization.md",7803,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fresource-organization.md",{"path":145,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":146,"sourceHtmlUrl":147},"references\u002Frubric.md",4162,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Frubric.md",{"path":149,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":150,"sourceHtmlUrl":151},"references\u002Fschemas.md",10007,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fschemas.md",{"path":153,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":154,"sourceHtmlUrl":155},"references\u002Fsecrets-and-credentials.md",18148,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fsecrets-and-credentials.md",{"path":157,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":158,"sourceHtmlUrl":159},"references\u002Fself-containment-principle.md",11362,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fself-containment-principle.md",{"path":161,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":162,"sourceHtmlUrl":163},"references\u002Fsize-limits.md",1555,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fsize-limits.md",{"path":165,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":166,"sourceHtmlUrl":167},"references\u002Fskill-categories.md",10936,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fskill-categories.md",{"path":169,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":170,"sourceHtmlUrl":171},"references\u002Fskill-template.md",1972,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fskill-template.md",{"path":173,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":174,"sourceHtmlUrl":175},"references\u002Fskill-workflow.md",18300,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fskill-workflow.md",{"path":177,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":178,"sourceHtmlUrl":179},"references\u002Fslash-command-conversion.md",15404,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fslash-command-conversion.md",{"path":181,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":182,"sourceHtmlUrl":183},"references\u002Ftask-management-patterns.md",3647,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Ftask-management-patterns.md",{"path":185,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":186,"sourceHtmlUrl":187},"references\u002Ftemplates.md",9205,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Ftemplates.md",{"path":189,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":190,"sourceHtmlUrl":191},"references\u002Ftessl-best-practices.md",2314,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Ftessl-best-practices.md",{"path":193,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":194,"sourceHtmlUrl":195},"references\u002Ftesting.md",4345,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Ftesting.md",{"path":197,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":198,"sourceHtmlUrl":199},"references\u002Ftroubleshooting-guide.md",6008,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Ftroubleshooting-guide.md",{"path":201,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":202,"sourceHtmlUrl":203},"references\u002Fweb-search-research.md",3830,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Freferences\u002Fweb-search-research.md",{"path":205,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":13,"sourceHtmlUrl":206},"scripts\u002F__init__.py","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002F__init__.py",{"path":208,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":209,"sourceHtmlUrl":210},"scripts\u002Faggregate_benchmark.py",16160,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Faggregate_benchmark.py",{"path":212,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":213,"sourceHtmlUrl":214},"scripts\u002Fanalyze_similarity.py",18591,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Fanalyze_similarity.py",{"path":216,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":217,"sourceHtmlUrl":218},"scripts\u002Fgenerate_report.py",12873,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Fgenerate_report.py",{"path":220,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":221,"sourceHtmlUrl":222},"scripts\u002Fimprove_description.py",15474,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Fimprove_description.py",{"path":224,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":225,"sourceHtmlUrl":226},"scripts\u002Finit_skill.py",11587,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Finit_skill.py",{"path":228,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":229,"sourceHtmlUrl":230},"scripts\u002Fpackage_skill.py",5114,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Fpackage_skill.py",{"path":232,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":233,"sourceHtmlUrl":234},"scripts\u002Fquick_validate.py",6189,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Fquick_validate.py",{"path":236,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":237,"sourceHtmlUrl":238},"scripts\u002Frepair_skill.py",8456,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Frepair_skill.py",{"path":240,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":241,"sourceHtmlUrl":242},"scripts\u002Frun_eval.py",13217,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Frun_eval.py",{"path":244,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":245,"sourceHtmlUrl":246},"scripts\u002Frun_loop.py",13951,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Frun_loop.py",{"path":248,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":249,"sourceHtmlUrl":250},"scripts\u002Fscan_skills.py",11261,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Fscan_skills.py",{"path":252,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":253,"sourceHtmlUrl":254},"scripts\u002Futils.py",4537,"https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fskill-development\u002Fscripts\u002Futils.py",[256,263,270,277,284,291],{"skillNo":257,"name":258,"nameZh":6,"description":259,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":260,"categories":261,"keywords":262,"publishedAt":19},"01M79R0K918Z0XDM74R0EY5H3P","demo","Fixture skill used by scripts.marketplace_ci sync tests.","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002Ftests\u002Fmarketplace_ci\u002Ffixtures\u002Fplugins\u002Fsample-kit\u002Fskills\u002Fdemo\u002FSKILL.md",[],[],{"skillNo":264,"name":265,"nameZh":6,"description":266,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":267,"categories":268,"keywords":269,"publishedAt":19},"01M79R08VMDJA2FPJKH3ZM309K","export-demo","Fixture skill used by scripts.marketplace_ci export tests.","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002Ftests\u002Fmarketplace_ci\u002Ffixtures\u002Fclaude\u002Fskills\u002Fexport-demo\u002FSKILL.md",[],[],{"skillNo":271,"name":272,"nameZh":6,"description":273,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":274,"categories":275,"keywords":276,"publishedAt":19},"01M79R042FS581DS2JNW5AZ1MK","work-to-development","Validate an accepted Linear Issue's readiness for implementation (criteria, dependencies, duplicate\u002Fexisting artifacts) and request governed branch\u002Fworktree creation through git-kit:starting-work — recording work-started Git\u002FGitHub evidence and a deliberate Linear Started transition only after Git identity is read back and verified. Use when asked to start work on a Linear issue, begin development for an accepted issue, or prepare an accepted issue for implementation. Never creates a branch directly — always delegates to git-kit.","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fworkmanagement-kit\u002Fskills\u002Fwork-to-development\u002FSKILL.md",[],[],{"skillNo":278,"name":279,"nameZh":6,"description":280,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":281,"categories":282,"keywords":283,"publishedAt":19},"01M79QZZD4EHB82P72P4N92BKT","work-linking","Create and maintain stable cross-system links and authority labels between Notion and Linear records, and detect\u002Frepair drift between them. Use when asked to link a Notion record to a Linear record, check for drift between Notion and Linear, or repair a broken\u002Fstale link. Repairs only the bounded non-authoritative copy or link — never chooses by newest timestamp, and never creates a reverse-write loop against either system's own authority.","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fworkmanagement-kit\u002Fskills\u002Fwork-linking\u002FSKILL.md",[],[],{"skillNo":285,"name":286,"nameZh":6,"description":287,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":288,"categories":289,"keywords":290,"publishedAt":19},"01M79QZVFA3GTVNZMJ51NESVSP","status-and-learning","Produce a dated, explicitly non-live Linear-to-Notion progress summary, or capture an outcome\u002Flearning after work completes. Use when asked to summarize progress to Notion, report portfolio status, or capture what was learned from a completed piece of work — not to disposition open follow-up items (see open-item-management for that). Never runs automatically or continuously — every summary is a deliberate, approved snapshot, not a live sync.","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fworkmanagement-kit\u002Fskills\u002Fstatus-and-learning\u002FSKILL.md",[],[],{"skillNo":292,"name":293,"nameZh":6,"description":294,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":295,"categories":296,"keywords":297,"publishedAt":19},"01M79QZNSKSABK7X2PT93X8QT5","repository-gates","Discover, delegate, record, and invalidate this repository's actual Git\u002FGitHub lifecycle gates (pre-commit, pre-push, PR required checks, Codex delta review, review requirements, merge rights) and the repository-policy provider profile that maps governed operations to git-kit. Use when asked what gates apply to this repository, to discover repository policy before starting Wave 2 work, or to check whether prior gate evidence is still valid after new commits. Discovery and delegation only — never executes a gate's own internal logic, and never invents a gate the repository doesn't actually define (in particular, never assumes a \"Review Changes\" gate exists).","https:\u002F\u002Fgithub.com\u002FAndreHahm\u002Fandres-cc-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fworkmanagement-kit\u002Fskills\u002Frepository-gates\u002FSKILL.md",[],[]]