[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M71SYDC8EBPAFJ7P9S7SCGV3":3,"skill-same:01M71SYDC8EBPAFJ7P9S7SCGV3":376},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":16,"sourceHtmlUrl":17,"categories":18,"keywords":35,"publishedAt":56,"linkOnly":57,"markdownContent":58,"markdownBody":59,"frontMatter":60,"refName":62,"commitSha":63,"skillDirectory":64,"forksCount":65,"repositoryHtmlUrl":66,"repositoryHomepageUrl":6,"ownerHtmlUrl":67,"ownerBlogUrl":6,"ownerDescription":68,"updatedAt":69,"install":70,"resources":83},"01M71SYDC8EBPAFJ7P9S7SCGV3","anthropic-grade-optimizer",null,"Audits and optimizes Claude-directing artifacts (CLAUDE.md, SKILL.md, subagent, hook, MCP config, system or user prompt, workflow, api_config) against 189 cited Anthropic rules across 11 dimensions, calibrated per model (Opus 4.7, Sonnet 4.6, Opus 4.6, Haiku 4.5). Every finding ships with verbatim quote and source URL; authorial voice is preserved.","l0z4n0-a1\u002Fskill-anthropic-grade-optimizer","l0z4n0-a1","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F220634244?v=4","01M71SYBEK8Q3WPCMCJ7A741A7","l0z4n0 🧙‍♂️ | João G. Lozano","01M71SYBFK0HAV1QD8MX39FTGW",20,"MIT","https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FSKILL.md",[19,24,27,30],{"categoryNo":20,"categoryName":21,"taxonomyType":22,"taxonomyTypeDesc":23},"01JBSEEDCAT000000000000007","代码审查","SCENARIO","使用场景",{"categoryNo":25,"categoryName":26,"taxonomyType":22,"taxonomyTypeDesc":23},"01JBSEEDCAT000000000000010","后端与接口",{"categoryNo":28,"categoryName":29,"taxonomyType":22,"taxonomyTypeDesc":23},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":31,"categoryName":32,"taxonomyType":33,"taxonomyTypeDesc":34},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55],"anthropic grade optimizer","optimizer","operator","mcp config","config","audit","artifact","haiku","optimize","across","review","dimensions","api","subagent","drift","workflow","URL","MCP","good","model","2026-11-10 13:02:02",false,"---\nname: anthropic-grade-optimizer\ndescription: Audits and optimizes Claude-directing artifacts (CLAUDE.md, SKILL.md, subagent, hook, MCP config, system or user prompt, workflow, api_config) against 189 cited Anthropic rules across 11 dimensions, calibrated per model (Opus 4.7, Sonnet 4.6, Opus 4.6, Haiku 4.5). Every finding ships with verbatim quote and source URL; authorial voice is preserved.\nwhen_to_use: Operator asks to audit, optimize, review, lint, score, refactor, or \"make Anthropic-grade\" any prompt, CLAUDE.md, SKILL.md, subagent, hook, MCP config, api_config, or workflow. Also triggers on model migration (\"good for Opus 4.7?\", \"optimize for Haiku\"), compliance review, or fixing adherence\u002Fverbosity\u002Fhallucination drift in a Claude artifact.\n---\n\n# Anthropic-Grade Optimizer\n\nAudits any Claude-directing artifact against the official Anthropic doctrine,\ncalibrates findings by target model, and proposes surgical optimizations that\npreserve authorial voice. Every finding cites a verbatim source URL; the skill\nships with cited rules only.\n\n## Contents\n\n- Hierarchy of authority\n- Honest scope\n- Quick start (and worked examples)\n- The Three Laws\n- Artifact types (9)\n- Target-model modulation\n- The 11 dimensions\n- Severity → triage\n- Emphasis conflict (D9, type-aware)\n- Detection methods\n- Operating modes\n- Adaptive modes (auto-trigger)\n- Output format (concise default)\n- Workflow (10 steps)\n- Scope discipline (positive framing)\n- Edge cases\n- Self-audit: Open Questions\n- References (10 files)\n- Assets (canonical snippets)\n- Scripts (6 entry points)\n- Validation (eval suite + strange-loop)\n\n## Hierarchy of authority\n\nAnthropic doctrine is the sole source of scoring rules. Optional interpretive\nlenses inform *how* to reason about findings (what to preserve, when to defer,\nhow to phrase recommendations) — they stay outside the rubric. On collision,\nAnthropic always wins.\n\n## Honest scope\n\n189 unique cited rules across 11 dimensions, each with a verbatim quote and\nsource URL. Read `references\u002Frules-anthropic.yaml` § `meta.total_unique_rules`\nfor the canonical count — every other counter in the skill must read from\nthere to prevent drift. Rules with deterministic detection (regex,\ncode-check) are audited automatically by `scripts\u002Fpass1_mechanical.py`. Rules\nwith qualitative criteria (llm-judge, heuristic) are audited by following\n`references\u002Fpass2-protocol.md`. Coverage gaps and known limitations live in\n`references\u002Fgaps.md` and must appear in the `coverage_caveat` block of every\nreport.\n\nPass-1 deterministic coverage in v1.2: ~52 rules (~28% of 189). Pass-2\ncovers the remaining ~137 qualitative rules. Hybrid detection is supported via\nthe `requires_pass2_grade` flag on findings.\n\n## Quick start\n\n**Canonical entry point (default):**\n\n```\npython scripts\u002Frun.py \u003Cartifact_path> --target opus-4-7 --mode audit\n```\n\nThe orchestrator chains classify → pass 1 → score → emit, and surfaces Pass 2\nprompts for the LLM. This is the single right answer for ~95% of audits.\n\n**Manual flow** is for one specific case: an operator forcing fine-grained\ncontrol over a single phase (re-running pass 1 only, scoring an external\nfindings file, debugging the classifier). Do not use it as a parallel path —\nthe orchestrator is the contract.\n\n1. **Ingest** — receive the artifact path or content, target model\n   (default `opus-4-7`), mode (default `audit`).\n2. **Classify** — Run `scripts\u002Fclassify_artifact.py` to detect artifact type\n   and load the rule subset from `references\u002Frubric-by-type.yaml`.\n3. **Audit** — Pass 1 mechanical via `scripts\u002Fpass1_mechanical.py`. Pass 2\n   qualitative reasoning following `references\u002Fpass2-protocol.md` against\n   `references\u002Frules-anthropic.yaml`.\n4. **Diagnose** — Classify each finding as 🔴 must-fix, 🟡 should-fix,\n   🟢 may-fix, ❓ open-question, or ⚪ preserve (authorial voice).\n5. **Optimize** *(when mode = optimize or full)* — Produce surgical diffs\n   per `references\u002Fpass2-protocol.md` § Diff Generation. Each diff cites\n   `source_url` + `verbatim_quote`. When a rule has a canonical snippet,\n   emit a verbatim patch from `assets\u002Fsnippets\u002F` rather than paraphrasing.\n6. **Validate** *(when mode = full)* — Re-score post-diff; abort when a\n   hard rule is introduced or voice drift exceeds 10%.\n7. **Emit** — Concise report (summary + scorecard + diff). Verbose mode\n   adds reasoning trail, preservation log, and open questions.\n\n### Worked example 1 — auditing a SKILL.md for Opus 4.7\n\n\u003Cexample>\nOperator request: \"audit this skill for Opus 4.7\"\nArtifact: `~\u002F.claude\u002Fskills\u002Fpdf-tools\u002FSKILL.md` (240 lines)\n\nStep 1 — Classify: type=`skill`, has_frontmatter=true, body_lines=235.\nStep 2 — Load rubric: 24 D-SKILL rules + 17 D-CLAR + 10 D-STRUCT + 8 D-EXAMPLE +\n         5 D-EVAL + safety. Suppress AR-CC-S09 doctrine-conflict?  No (skill body).\nStep 3 — Pass 1 fires: AR-CC-S20 (lib mentioned without `pip install`), AR-CC-S22\n         (3 script refs unframed), AR-CLAR-006 (7 negatives, 1 positive alt).\nStep 4 — Triage: 1 🔴 (AR-CC-S20), 2 🟢 (S22, CLAR-006).\nStep 5 — Optimize emits snippet patches: none (no rule with canonical snippet\n         fires). Inline diff for AR-CC-S20 adds \"Install required package: pip install pypdf\".\nStep 6 — Validate: post-diff score 92 (was 78). Voice drift 4% — under 10% gate.\nStep 7 — Emit concise report.\n\u003C\u002Fexample>\n\n### Worked example 2 — strange-loop self-audit\n\n\u003Cexample>\nOperator request: \"run the skill on its own SKILL.md\"\nArtifact: `anthropic-grade-optimizer\u002FSKILL.md`\n\nStep 3 — Pass 1 fires: AR-CC-S14 (name contains \"anthropic\" reserved word) and\n         possibly AR-CC-S21 (TOC) and AR-CC-S22 (script framing).\nStep 4 — AR-CC-S14: see § Self-audit: Open Questions below for the operator's\n         documented decision (semantic-justification exception).\nStep 5 — No-op for the AR-CC-S14 finding (declared exception).\nStep 7 — Concise report flags the open question and links to the §.\n\u003C\u002Fexample>\n\n### Worked example 3 — auditing an api_config snippet\n\n\u003Cexample>\nOperator request: \"is this Python snippet safe for Opus 4.7?\"\nArtifact: a `client.messages.create(...)` call with `temperature=0.7`,\n          `effort='low'`, last message role=assistant.\nType: `api_config`. Target: `claude-opus-4-7`.\n\nStep 3 — Pass 1 fires: AR-MODEL-002 prefill (HARD, last role=assistant);\n         AP-15 sampling param `temperature` (HARD); AR-REASON-017 effort=low\n         on opus-4-7 with coding signal (severity_amplification → HARD);\n         AR-MODEL-021..025 emitted as one Open Question with 5 options\n         (collapsed via `open_question=True`).\nStep 4 — Triage: 3 🔴 (002, AP-15, REASON-017), 1 ❓ (021..025).\nStep 5 — Optimize emits inline diffs for AR-MODEL-002 \u002F AP-15 (remove sampling\n         params, move continuation to user message per AR-MODEL-024 if that\n         pattern is the operator's choice).\n\u003C\u002Fexample>\n\n## The Three Laws\n\nThese three laws encode the discipline that separates Anthropic-grade from\n\"looks rigorous\":\n\n1. **Cite or stay silent.** Every 🔴 \u002F 🟡 finding carries a `source_url`.\n   When a source is absent, the finding is downgraded to `EXTERNAL_ENRICHMENT`\n   or dropped — Anthropic-grade ships only cited rules.\n2. **Artifact type comes first.** Apply the rule subset for the detected\n   type. Firing a SKILL.md rule against a CLAUDE.md is a false positive — see\n   `references\u002Frubric-by-type.yaml` § `false_positive_rules`.\n3. **Voice drift trumps score.** Raising the score by diluting the\n   operator's voice is a regression in disguise. Optimizations with\n   `voice_drift > 10%` abort; with `--push-ceiling` the gate tightens to 5%.\n\n## Artifact types (9)\n\nDetection happens on filename plus content; each type loads a tailored rule\nsubset from `references\u002Frubric-by-type.yaml`:\n\n| Type | Path signal | Primary dimensions |\n|---|---|---|\n| `claude_md` | `CLAUDE.md`, `CLAUDE.local.md` | D-CC (memory), D-CLAR, D-STRUCT |\n| `skill` | `SKILL.md` in `skills\u002F\u003Cname>\u002F` | D-CC (skill), D-CLAR, D-STRUCT |\n| `slash_command` | `.claude\u002Fcommands\u002F\u003Cname>.md` (legacy) | D-CC (skill subset) |\n| `subagent` | `.claude\u002Fagents\u002F\u003Cname>.md` | D-CC (subagent), D-AGENT, D-CLAR |\n| `hook_config` | `settings.json` `hooks` key | D-CC (hooks) |\n| `mcp_config` | `.mcp.json`, `.claude.json` | D-CC (mcp) |\n| `system_prompt` \u002F `user_prompt` | inline \u002F API artifact | D-CLAR, D-STRUCT, D-EXAMPLE, D-REASON, D-CONTEXT, D-MODEL, D-TOOL, D-VISION |\n| `api_config` | Python\u002FJSON snippet with `client.messages.create` or `model=\"claude-...\"` | D-MODEL, D-REASON, D-CONTEXT (cache), D-TOOL, D-VISION |\n| `workflow` | YAML pipeline | D-AGENT, D-EVAL |\n\n## Target-model modulation\n\nEach model has a profile cell in `references\u002Fmodulation-matrix.yaml`. Critical\nanti-patterns flagged automatically:\n\n- **Opus 4.7** — rejects `temperature` \u002F `top_p` \u002F `top_k` (HTTP 400),\n  rejects manual `thinking:{enabled,budget_tokens}` (400), rejects prefill\n  on the last assistant turn (400). Interprets prompts literally — state\n  scope explicitly. ALL-CAPS imperatives over-suppress output. Effort levels\n  are strict; pair `xhigh` with coding\u002Fagentic. Tool use decreases by\n  default; raise effort or prompt explicitly when more tool use is wanted.\n- **Opus 4.6 \u002F Sonnet 4.6** — rejects prefill (400). Adaptive thinking is\n  recommended.\n- **Haiku 4.5** — adaptive thinking unsupported, `effort` parameter\n  unsupported, 200k-token context. Needs maximum explicitness; few-shot is\n  more critical (5+ examples).\n\nWhen a target model lacks a profile (e.g., a new release), inherit from the\nclosest sibling and stamp the report with `confidence: low`. Fallback rules\nlive in `references\u002Fmodulation-matrix.yaml`.\n\n## The 11 dimensions\n\n| Dim | Focus | Weight |\n|---|---|---|\n| D-CLAR | Clarity, directness, positive framing | 16% |\n| D-STRUCT | XML tags, hierarchy, formatting | 11% |\n| D-EXAMPLE | Few-shot 3-5 examples in `\u003Cexample>` | 8% |\n| D-REASON | CoT, adaptive thinking, self-check | 12% |\n| D-CONTEXT | Doc placement, grounding, cache | 10% |\n| D-MODEL | Model-specific calibration | 10% |\n| D-AGENT | Stop conditions, ground truth, delegation | 8% |\n| D-EVAL | Verification embedded in prompt | 7% |\n| D-TOOL | Tool definitions, schemas, parallel | 6% |\n| D-VISION | Image placement, coords, formats | 4% |\n| D-CC | Claude Code ecosystem (memory\u002Fskill\u002Fagent\u002Fhook\u002FMCP) | 8% |\n\n## Severity → triage\n\n| Severity | Score impact | Triage |\n|---|---|---|\n| `hard` \u002F `critical` \u002F `never` | -10 to -20 | 🔴 must-fix |\n| `soft` \u002F `high` | -5 to -8 | 🟡 should-fix |\n| `medium` | -3 to -5 | 🟡 should-fix |\n| `low` \u002F `warning` | -1 to -2 | 🟢 may-fix |\n| (5+ migration options, group fix) | n\u002Fa | ❓ open-question |\n| (authorial voice signal) | n\u002Fa | ⚪ preserve |\n\n## Emphasis conflict (D9, type-aware)\n\nTwo Anthropic sources disagree about ALL-CAPS \u002F `IMPORTANT` \u002F `YOU MUST`:\n\n- `code.claude.com\u002Fdocs\u002Fen\u002Fbest-practices` endorses emphasis to tune adherence.\n- `github.com\u002Fanthropics\u002Fskills\u002Fskill-creator` flags ALL-CAPS as yellow,\n  preferring an explanation of the underlying reason.\n\nResolution applies by artifact type:\n\n| Artifact type | Emphasis verdict |\n|---|---|\n| `claude_md`, `system_prompt`, `user_prompt` | Endorsed (suppress yellow-flag finding) |\n| `skill`, `subagent` body | Yellow flag (raise finding, explain WHY) |\n| `tool_description` | Neutral (no doctrine) |\n\nWhen AR-CC-S09 fires on a `claude_md`, suppress it as a false positive\n(anti-pattern AP-10).\n\n## Detection methods\n\nPrefer determinism (cheapest instrument that solves):\n\n- **regex** \u002F **code-check** → `scripts\u002Fpass1_mechanical.py`\n- **heuristic** (line counts, structural shape) → `scripts\u002Fpass1_mechanical.py`\n- **hybrid** (regex emits candidate, LLM grades substance) → Pass 1 emits\n  with `requires_pass2_grade=true`; Pass 2 finalizes\n- **llm-judge** → reason directly with the artifact text loaded; cite the\n  rule + verbatim quote\n\nReach for `llm-judge` only when no script can decide.\n\n## Operating modes\n\n- `audit` *(default)* — read-only report, no diffs proposed.\n- `optimize` — generate diffs without validating.\n- `full` — audit → optimize → validate → emit. Loops up to three passes\n  when score \u003C 85.\n- `--push-ceiling` — applies may-fix (🟢) findings; tightens voice_drift to 5%.\n- `--fast` — top-5 findings only; single pass.\n- `--no-exceptions` — disables operator-declared exceptions (see § Self-audit).\n  Reports the strict raw score instead of the active score. The orchestrator\n  always shows both side by side when an exception is in effect, so this flag\n  is for downstream tooling that needs the unadjusted number.\n\n## Adaptive modes (auto-trigger)\n\n| Trigger | Mode | Effect |\n|---|---|---|\n| Artifact > 800 lines | STRATIFIED-AUDIT | Map blocks → triage → deep-audit top-K (default K=5) |\n| Initial score ≥ 85 + zero 🔴 | AFFIRMATION | Short report; surface 🟢 only with `--push-ceiling` |\n| Artifact \u003C 50 tokens + minimal markers | HIGH-UNCERTAINTY | Ask \"intentional brevity or unfinished?\" before expanding |\n| Target model lacks profile | FALLBACK-MODEL | Inherit sibling, mark `INFERRED`, reduce D-MODEL weight 10% → 5% |\n\n## Output format (concise default)\n\n```\nANTHROPIC-GRADE OPTIMIZER REPORT\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\nArtifact: \u003Cpath>\nType: \u003Ctype> | Target: \u003Cmodel> | Mode: \u003Cmode> | Rules: v1.2 (189)\n\nSCORE: \u003C0-100>  (\u003Cgrade>)\n  D-CLAR: \u003Cn>%   D-STRUCT: \u003Cn>%   D-EXAMPLE: \u003Cn>%\n  D-REASON: \u003Cn>% D-CONTEXT: \u003Cn>%  D-MODEL: \u003Cn>%\n  D-AGENT: \u003Cn>%  D-EVAL: \u003Cn>%     D-TOOL: \u003Cn>%\n  D-VISION: \u003Cn>% D-CC: \u003Cn>%\n\nFINDINGS: \u003Ccount_red> 🔴 \u002F \u003Ccount_yellow> 🟡 \u002F \u003Ccount_green> 🟢 \u002F \u003Ccount_q> ❓ \u002F \u003Ccount_preserve> ⚪\n\n🔴 Must-fix\n  [F-001] \u003Crule_id> — \u003Cone-line statement>\n    Location: \u003Cline range>\n    Source: \u003Csource_url>\n    Quote: \"\u003Cverbatim>\"\n\n🟡 Should-fix\n  ...\n\n❓ Open questions (operator decides)\n  [Q-001] \u003Crule_id> — \u003Coptions summary>\n\nDIFF (mode ≠ audit):\n  \u003Cunified diff with rule citations>\nSNIPPET PATCHES (when applicable):\n  \u003Cverbatim insertions from assets\u002Fsnippets\u002F with source_url>\n\nCOVERAGE CAVEAT:\n  Doctrine dissected on 2026-05-02. 189 cited rules. Gaps in references\u002Fgaps.md.\n  \u003Cif model is INFERRED>: Target model profile inherited; confidence: low.\n```\n\nVerbose adds: per-finding reasoning trail, preservation log (what stayed and\nwhy), and explicit open questions for authorial decisions.\n\n## Workflow\n\n1. Run `python scripts\u002Fclassify_artifact.py \u003Cpath>` → `{type, line_count, signals}`.\n2. Read `references\u002Frubric-by-type.yaml` for the type's rule_ids.\n3. Run `python scripts\u002Fpass1_mechanical.py \u003Cpath> --type \u003Ctype> --target-model \u003Cmodel>`\n   → deterministic findings (including hybrid candidates).\n4. For each rule with `detection_method=llm-judge` in the type's subset,\n   read the relevant section, reason about adherence, cite the rule's\n   `verbatim_quote`, and append the finding to the bundle.\n5. Apply D9 emphasis-conflict resolution by artifact type\n   (see `references\u002Frubric-by-type.yaml` § `false_positive_rules`).\n6. Run `python scripts\u002Fscore_calculator.py \u003Cfindings.json>`\n   → `{score, per_dimension, voice_drift_estimate}`.\n7. Triage findings (🔴\u002F🟡\u002F🟢\u002F❓\u002F⚪).\n8. When mode ≠ `audit`, run `python scripts\u002Foptimize.py \u003Cpath> --findings \u003Cjson> --type \u003Ctype>`\n   to produce surgical diffs and snippet patches (each cited).\n9. When mode = `full`, simulate post-diff, re-score, validate gate.\n10. Emit report (concise default; verbose with `--verbose`).\n\n## Scope discipline (positive framing)\n\nThis skill audits prompts and Claude-directing artifacts. Its responsibilities:\n\n- **Audits**: cited findings only, with `verbatim_quote` + `source_url`.\n- **Proposes diffs**: surgical, one rule per diff, voice-preserving.\n- **Stays in source language**: preserves the artifact's authorial voice.\n- **Defers to operator**: the operator stays in the editor's seat — diffs\n  apply only with `--apply` to an `\u003Cpath>.optimized.\u003Cext>` sibling, never\n  overwriting the original.\n\nOut-of-scope (sibling tools handle these):\n\n- Evaluating Claude's *responses* — that work belongs to eval pipelines.\n- Auditing application code that *consumes* the Claude API beyond the\n  prompts and artifacts themselves.\n- Inventing rules without a source URL — every finding ships cited.\n\n## Edge cases\n\nWhen a rule could fire but no verbatim source quote can be found, hold the\nfinding back. Add it to the report's `coverage_caveat` block as a candidate\nfor the next dissection round. Anthropic-grade means cited or silent.\n\nWhen two rules conflict and no documented resolution exists (see\n`references\u002Fgaps.md` for the 10 known conflicts), surface both findings and\nelevate the decision to the **open questions** block — the operator decides.\n\n## Self-audit: Open Questions\n\nThe skill audits itself on every release via the canonical entry point:\n\n```\npython scripts\u002Frun.py SKILL.md --target opus-4-7 --mode audit\n```\n\nDeclared exceptions (operator-justified, registered in `scripts\u002Frun.py § DECLARED_EXCEPTIONS`):\n\n- **AR-CC-S14 — reserved word \"anthropic\" in name.** The doctrine quote\n  *\"Cannot contain reserved words: 'anthropic', 'claude'.\"* applies to skill\n  names. The operator decision is to keep the name\n  `anthropic-grade-optimizer` because it is the most precise descriptor of\n  the skill's purpose. A rename to `doctrine-grade-optimizer` would lose the\n  explicit doctrinal anchor and degrade trigger reliability for operators\n  searching by domain. The Pass-1 detector continues to flag it on every\n  run; the orchestrator splits the report into `active score` (exception\n  applied) and `raw score` (`--no-exceptions`) so consumers retain full\n  information. D-CC dimension shows the residual penalty\n  transparently — voice fidelity over hidden adjustment.\n\nThe full self-audit run, including findings, both scores, and dimension\nbreakdown, is committed at `evals\u002FSELF-AUDIT.md` and re-runs on every release.\n\n## References (10 files in `references\u002F`)\n\n- `rules-anthropic.yaml` — 189 rules with `source_url` + `verbatim_quote` +\n  `detection_method` (v1.2). The canonical SSOT for rule count and content.\n- `rules-master.yaml` — index by dimension (points into rules-anthropic.yaml).\n- `modulation-matrix.yaml` — 4 models × ~10 cells, each\n  `{value, source_url, confidence}`.\n- `rubric-by-type.yaml` — rule subsets per artifact type, with\n  false-positive suppression matrix and 3 documented doctrinal conflicts.\n- `quotes-canonical.md` — verbatim citations for report output.\n- `heuristics.md` — H1-H12 operational shortcuts.\n- `anti-patterns.md` — AP-1 to AP-15 with detection method.\n- `gaps.md` — known doctrinal gaps (declare in every report).\n- `pass2-protocol.md` — qualitative auditing workflow (llm-judge rules).\n\n## Assets (canonical snippets)\n\n- `assets\u002Fsnippets\u002F` — 18 canonical snippets (16 rule-bound + 2 advisory).\n  Every snippet is byte-by-byte verbatim from the live Anthropic doc, with\n  a header table declaring `triggers_rule`, `applies_to`, `model_specific`,\n  `source_url`, `insertion_pattern`, and `voice_drift`. Snippets cover\n  high-leverage rules: AR-MODEL-018 (code-review coverage), AR-MODEL-011\n  (investigate before answering), AR-MODEL-007\u002FAR-AGENT-009 (parallel\n  tool calls), AR-CLAR-011 (action vs research posture, two snippets),\n  and 12 more.\n- `assets\u002Fsnippets\u002Findex.yaml` — `rule_id → snippet` map consumed by\n  `scripts\u002Foptimize.py`. Type-aware suppression prevents AP-9 false patches\n  (e.g., subagent files never receive subagent-orchestration snippets).\n\n## Scripts (6 entry points in `scripts\u002F`)\n\n- `run.py` — orchestrator; runs classify → pass 1 → score → emit. Single\n  entry point for most workflows.\n- `classify_artifact.py \u003Cpath>` — returns\n  `{type, line_count, language, has_frontmatter, signals}`.\n- `pass1_mechanical.py \u003Cpath> --type \u003Ctype> --target-model \u003Cmodel>` —\n  emits findings array (regex + code-check + hybrid candidates).\n- `score_calculator.py \u003Cfindings.json>` — emits\n  `{score, per_dimension, voice_drift_estimate}`.\n- `optimize.py \u003Cpath> --findings \u003Cjson> --type \u003Ctype> [--apply]` — emits\n  diffs and snippet patches; with `--apply` writes to\n  `\u003Cpath>.optimized.\u003Cext>` (preserving the original).\n- `run_eval_suite.py` — replays `evals\u002Ffixtures\u002F` ground-truth suite.\n\nAll scripts use forward slashes and read paths via the `CLAUDE_SKILL_DIR`\nenv var, defaulting to the parent of the script directory (per AR-CC-S06).\nOn Windows, scripts force UTF-8 stdout to render the triage glyphs\n(🔴\u002F🟡\u002F🟢\u002F❓\u002F⚪) without `cp1252` crashes.\n\n## Validation (eval suite + strange-loop)\n\nThe skill validates itself via `evals\u002F`. The fixture set covers good \u002F\nbad \u002F adversarial-minimal cases for every artifact type, plus cross-type\nfixtures that exercise the false-positive suppression matrix. Ground-truth\nscores live in `evals\u002Fground-truth.yaml`; `scripts\u002Frun_eval_suite.py`\nreplays them. The strange-loop self-audit\n(`evals\u002FSELF-AUDIT.md`) re-runs the skill against its own SKILL.md on every\nrelease and commits the result for transparency.\n","# Anthropic-Grade Optimizer\n\nAudits any Claude-directing artifact against the official Anthropic doctrine,\ncalibrates findings by target model, and proposes surgical optimizations that\npreserve authorial voice. Every finding cites a verbatim source URL; the skill\nships with cited rules only.\n\n## Contents\n\n- Hierarchy of authority\n- Honest scope\n- Quick start (and worked examples)\n- The Three Laws\n- Artifact types (9)\n- Target-model modulation\n- The 11 dimensions\n- Severity → triage\n- Emphasis conflict (D9, type-aware)\n- Detection methods\n- Operating modes\n- Adaptive modes (auto-trigger)\n- Output format (concise default)\n- Workflow (10 steps)\n- Scope discipline (positive framing)\n- Edge cases\n- Self-audit: Open Questions\n- References (10 files)\n- Assets (canonical snippets)\n- Scripts (6 entry points)\n- Validation (eval suite + strange-loop)\n\n## Hierarchy of authority\n\nAnthropic doctrine is the sole source of scoring rules. Optional interpretive\nlenses inform *how* to reason about findings (what to preserve, when to defer,\nhow to phrase recommendations) — they stay outside the rubric. On collision,\nAnthropic always wins.\n\n## Honest scope\n\n189 unique cited rules across 11 dimensions, each with a verbatim quote and\nsource URL. Read `references\u002Frules-anthropic.yaml` § `meta.total_unique_rules`\nfor the canonical count — every other counter in the skill must read from\nthere to prevent drift. Rules with deterministic detection (regex,\ncode-check) are audited automatically by `scripts\u002Fpass1_mechanical.py`. Rules\nwith qualitative criteria (llm-judge, heuristic) are audited by following\n`references\u002Fpass2-protocol.md`. Coverage gaps and known limitations live in\n`references\u002Fgaps.md` and must appear in the `coverage_caveat` block of every\nreport.\n\nPass-1 deterministic coverage in v1.2: ~52 rules (~28% of 189). Pass-2\ncovers the remaining ~137 qualitative rules. Hybrid detection is supported via\nthe `requires_pass2_grade` flag on findings.\n\n## Quick start\n\n**Canonical entry point (default):**\n\n```\npython scripts\u002Frun.py \u003Cartifact_path> --target opus-4-7 --mode audit\n```\n\nThe orchestrator chains classify → pass 1 → score → emit, and surfaces Pass 2\nprompts for the LLM. This is the single right answer for ~95% of audits.\n\n**Manual flow** is for one specific case: an operator forcing fine-grained\ncontrol over a single phase (re-running pass 1 only, scoring an external\nfindings file, debugging the classifier). Do not use it as a parallel path —\nthe orchestrator is the contract.\n\n1. **Ingest** — receive the artifact path or content, target model\n   (default `opus-4-7`), mode (default `audit`).\n2. **Classify** — Run `scripts\u002Fclassify_artifact.py` to detect artifact type\n   and load the rule subset from `references\u002Frubric-by-type.yaml`.\n3. **Audit** — Pass 1 mechanical via `scripts\u002Fpass1_mechanical.py`. Pass 2\n   qualitative reasoning following `references\u002Fpass2-protocol.md` against\n   `references\u002Frules-anthropic.yaml`.\n4. **Diagnose** — Classify each finding as 🔴 must-fix, 🟡 should-fix,\n   🟢 may-fix, ❓ open-question, or ⚪ preserve (authorial voice).\n5. **Optimize** *(when mode = optimize or full)* — Produce surgical diffs\n   per `references\u002Fpass2-protocol.md` § Diff Generation. Each diff cites\n   `source_url` + `verbatim_quote`. When a rule has a canonical snippet,\n   emit a verbatim patch from `assets\u002Fsnippets\u002F` rather than paraphrasing.\n6. **Validate** *(when mode = full)* — Re-score post-diff; abort when a\n   hard rule is introduced or voice drift exceeds 10%.\n7. **Emit** — Concise report (summary + scorecard + diff). Verbose mode\n   adds reasoning trail, preservation log, and open questions.\n\n### Worked example 1 — auditing a SKILL.md for Opus 4.7\n\n\u003Cexample>\nOperator request: \"audit this skill for Opus 4.7\"\nArtifact: `~\u002F.claude\u002Fskills\u002Fpdf-tools\u002FSKILL.md` (240 lines)\n\nStep 1 — Classify: type=`skill`, has_frontmatter=true, body_lines=235.\nStep 2 — Load rubric: 24 D-SKILL rules + 17 D-CLAR + 10 D-STRUCT + 8 D-EXAMPLE +\n         5 D-EVAL + safety. Suppress AR-CC-S09 doctrine-conflict?  No (skill body).\nStep 3 — Pass 1 fires: AR-CC-S20 (lib mentioned without `pip install`), AR-CC-S22\n         (3 script refs unframed), AR-CLAR-006 (7 negatives, 1 positive alt).\nStep 4 — Triage: 1 🔴 (AR-CC-S20), 2 🟢 (S22, CLAR-006).\nStep 5 — Optimize emits snippet patches: none (no rule with canonical snippet\n         fires). Inline diff for AR-CC-S20 adds \"Install required package: pip install pypdf\".\nStep 6 — Validate: post-diff score 92 (was 78). Voice drift 4% — under 10% gate.\nStep 7 — Emit concise report.\n\u003C\u002Fexample>\n\n### Worked example 2 — strange-loop self-audit\n\n\u003Cexample>\nOperator request: \"run the skill on its own SKILL.md\"\nArtifact: `anthropic-grade-optimizer\u002FSKILL.md`\n\nStep 3 — Pass 1 fires: AR-CC-S14 (name contains \"anthropic\" reserved word) and\n         possibly AR-CC-S21 (TOC) and AR-CC-S22 (script framing).\nStep 4 — AR-CC-S14: see § Self-audit: Open Questions below for the operator's\n         documented decision (semantic-justification exception).\nStep 5 — No-op for the AR-CC-S14 finding (declared exception).\nStep 7 — Concise report flags the open question and links to the §.\n\u003C\u002Fexample>\n\n### Worked example 3 — auditing an api_config snippet\n\n\u003Cexample>\nOperator request: \"is this Python snippet safe for Opus 4.7?\"\nArtifact: a `client.messages.create(...)` call with `temperature=0.7`,\n          `effort='low'`, last message role=assistant.\nType: `api_config`. Target: `claude-opus-4-7`.\n\nStep 3 — Pass 1 fires: AR-MODEL-002 prefill (HARD, last role=assistant);\n         AP-15 sampling param `temperature` (HARD); AR-REASON-017 effort=low\n         on opus-4-7 with coding signal (severity_amplification → HARD);\n         AR-MODEL-021..025 emitted as one Open Question with 5 options\n         (collapsed via `open_question=True`).\nStep 4 — Triage: 3 🔴 (002, AP-15, REASON-017), 1 ❓ (021..025).\nStep 5 — Optimize emits inline diffs for AR-MODEL-002 \u002F AP-15 (remove sampling\n         params, move continuation to user message per AR-MODEL-024 if that\n         pattern is the operator's choice).\n\u003C\u002Fexample>\n\n## The Three Laws\n\nThese three laws encode the discipline that separates Anthropic-grade from\n\"looks rigorous\":\n\n1. **Cite or stay silent.** Every 🔴 \u002F 🟡 finding carries a `source_url`.\n   When a source is absent, the finding is downgraded to `EXTERNAL_ENRICHMENT`\n   or dropped — Anthropic-grade ships only cited rules.\n2. **Artifact type comes first.** Apply the rule subset for the detected\n   type. Firing a SKILL.md rule against a CLAUDE.md is a false positive — see\n   `references\u002Frubric-by-type.yaml` § `false_positive_rules`.\n3. **Voice drift trumps score.** Raising the score by diluting the\n   operator's voice is a regression in disguise. Optimizations with\n   `voice_drift > 10%` abort; with `--push-ceiling` the gate tightens to 5%.\n\n## Artifact types (9)\n\nDetection happens on filename plus content; each type loads a tailored rule\nsubset from `references\u002Frubric-by-type.yaml`:\n\n| Type | Path signal | Primary dimensions |\n|---|---|---|\n| `claude_md` | `CLAUDE.md`, `CLAUDE.local.md` | D-CC (memory), D-CLAR, D-STRUCT |\n| `skill` | `SKILL.md` in `skills\u002F\u003Cname>\u002F` | D-CC (skill), D-CLAR, D-STRUCT |\n| `slash_command` | `.claude\u002Fcommands\u002F\u003Cname>.md` (legacy) | D-CC (skill subset) |\n| `subagent` | `.claude\u002Fagents\u002F\u003Cname>.md` | D-CC (subagent), D-AGENT, D-CLAR |\n| `hook_config` | `settings.json` `hooks` key | D-CC (hooks) |\n| `mcp_config` | `.mcp.json`, `.claude.json` | D-CC (mcp) |\n| `system_prompt` \u002F `user_prompt` | inline \u002F API artifact | D-CLAR, D-STRUCT, D-EXAMPLE, D-REASON, D-CONTEXT, D-MODEL, D-TOOL, D-VISION |\n| `api_config` | Python\u002FJSON snippet with `client.messages.create` or `model=\"claude-...\"` | D-MODEL, D-REASON, D-CONTEXT (cache), D-TOOL, D-VISION |\n| `workflow` | YAML pipeline | D-AGENT, D-EVAL |\n\n## Target-model modulation\n\nEach model has a profile cell in `references\u002Fmodulation-matrix.yaml`. Critical\nanti-patterns flagged automatically:\n\n- **Opus 4.7** — rejects `temperature` \u002F `top_p` \u002F `top_k` (HTTP 400),\n  rejects manual `thinking:{enabled,budget_tokens}` (400), rejects prefill\n  on the last assistant turn (400). Interprets prompts literally — state\n  scope explicitly. ALL-CAPS imperatives over-suppress output. Effort levels\n  are strict; pair `xhigh` with coding\u002Fagentic. Tool use decreases by\n  default; raise effort or prompt explicitly when more tool use is wanted.\n- **Opus 4.6 \u002F Sonnet 4.6** — rejects prefill (400). Adaptive thinking is\n  recommended.\n- **Haiku 4.5** — adaptive thinking unsupported, `effort` parameter\n  unsupported, 200k-token context. Needs maximum explicitness; few-shot is\n  more critical (5+ examples).\n\nWhen a target model lacks a profile (e.g., a new release), inherit from the\nclosest sibling and stamp the report with `confidence: low`. Fallback rules\nlive in `references\u002Fmodulation-matrix.yaml`.\n\n## The 11 dimensions\n\n| Dim | Focus | Weight |\n|---|---|---|\n| D-CLAR | Clarity, directness, positive framing | 16% |\n| D-STRUCT | XML tags, hierarchy, formatting | 11% |\n| D-EXAMPLE | Few-shot 3-5 examples in `\u003Cexample>` | 8% |\n| D-REASON | CoT, adaptive thinking, self-check | 12% |\n| D-CONTEXT | Doc placement, grounding, cache | 10% |\n| D-MODEL | Model-specific calibration | 10% |\n| D-AGENT | Stop conditions, ground truth, delegation | 8% |\n| D-EVAL | Verification embedded in prompt | 7% |\n| D-TOOL | Tool definitions, schemas, parallel | 6% |\n| D-VISION | Image placement, coords, formats | 4% |\n| D-CC | Claude Code ecosystem (memory\u002Fskill\u002Fagent\u002Fhook\u002FMCP) | 8% |\n\n## Severity → triage\n\n| Severity | Score impact | Triage |\n|---|---|---|\n| `hard` \u002F `critical` \u002F `never` | -10 to -20 | 🔴 must-fix |\n| `soft` \u002F `high` | -5 to -8 | 🟡 should-fix |\n| `medium` | -3 to -5 | 🟡 should-fix |\n| `low` \u002F `warning` | -1 to -2 | 🟢 may-fix |\n| (5+ migration options, group fix) | n\u002Fa | ❓ open-question |\n| (authorial voice signal) | n\u002Fa | ⚪ preserve |\n\n## Emphasis conflict (D9, type-aware)\n\nTwo Anthropic sources disagree about ALL-CAPS \u002F `IMPORTANT` \u002F `YOU MUST`:\n\n- `code.claude.com\u002Fdocs\u002Fen\u002Fbest-practices` endorses emphasis to tune adherence.\n- `github.com\u002Fanthropics\u002Fskills\u002Fskill-creator` flags ALL-CAPS as yellow,\n  preferring an explanation of the underlying reason.\n\nResolution applies by artifact type:\n\n| Artifact type | Emphasis verdict |\n|---|---|\n| `claude_md`, `system_prompt`, `user_prompt` | Endorsed (suppress yellow-flag finding) |\n| `skill`, `subagent` body | Yellow flag (raise finding, explain WHY) |\n| `tool_description` | Neutral (no doctrine) |\n\nWhen AR-CC-S09 fires on a `claude_md`, suppress it as a false positive\n(anti-pattern AP-10).\n\n## Detection methods\n\nPrefer determinism (cheapest instrument that solves):\n\n- **regex** \u002F **code-check** → `scripts\u002Fpass1_mechanical.py`\n- **heuristic** (line counts, structural shape) → `scripts\u002Fpass1_mechanical.py`\n- **hybrid** (regex emits candidate, LLM grades substance) → Pass 1 emits\n  with `requires_pass2_grade=true`; Pass 2 finalizes\n- **llm-judge** → reason directly with the artifact text loaded; cite the\n  rule + verbatim quote\n\nReach for `llm-judge` only when no script can decide.\n\n## Operating modes\n\n- `audit` *(default)* — read-only report, no diffs proposed.\n- `optimize` — generate diffs without validating.\n- `full` — audit → optimize → validate → emit. Loops up to three passes\n  when score \u003C 85.\n- `--push-ceiling` — applies may-fix (🟢) findings; tightens voice_drift to 5%.\n- `--fast` — top-5 findings only; single pass.\n- `--no-exceptions` — disables operator-declared exceptions (see § Self-audit).\n  Reports the strict raw score instead of the active score. The orchestrator\n  always shows both side by side when an exception is in effect, so this flag\n  is for downstream tooling that needs the unadjusted number.\n\n## Adaptive modes (auto-trigger)\n\n| Trigger | Mode | Effect |\n|---|---|---|\n| Artifact > 800 lines | STRATIFIED-AUDIT | Map blocks → triage → deep-audit top-K (default K=5) |\n| Initial score ≥ 85 + zero 🔴 | AFFIRMATION | Short report; surface 🟢 only with `--push-ceiling` |\n| Artifact \u003C 50 tokens + minimal markers | HIGH-UNCERTAINTY | Ask \"intentional brevity or unfinished?\" before expanding |\n| Target model lacks profile | FALLBACK-MODEL | Inherit sibling, mark `INFERRED`, reduce D-MODEL weight 10% → 5% |\n\n## Output format (concise default)\n\n```\nANTHROPIC-GRADE OPTIMIZER REPORT\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\nArtifact: \u003Cpath>\nType: \u003Ctype> | Target: \u003Cmodel> | Mode: \u003Cmode> | Rules: v1.2 (189)\n\nSCORE: \u003C0-100>  (\u003Cgrade>)\n  D-CLAR: \u003Cn>%   D-STRUCT: \u003Cn>%   D-EXAMPLE: \u003Cn>%\n  D-REASON: \u003Cn>% D-CONTEXT: \u003Cn>%  D-MODEL: \u003Cn>%\n  D-AGENT: \u003Cn>%  D-EVAL: \u003Cn>%     D-TOOL: \u003Cn>%\n  D-VISION: \u003Cn>% D-CC: \u003Cn>%\n\nFINDINGS: \u003Ccount_red> 🔴 \u002F \u003Ccount_yellow> 🟡 \u002F \u003Ccount_green> 🟢 \u002F \u003Ccount_q> ❓ \u002F \u003Ccount_preserve> ⚪\n\n🔴 Must-fix\n  [F-001] \u003Crule_id> — \u003Cone-line statement>\n    Location: \u003Cline range>\n    Source: \u003Csource_url>\n    Quote: \"\u003Cverbatim>\"\n\n🟡 Should-fix\n  ...\n\n❓ Open questions (operator decides)\n  [Q-001] \u003Crule_id> — \u003Coptions summary>\n\nDIFF (mode ≠ audit):\n  \u003Cunified diff with rule citations>\nSNIPPET PATCHES (when applicable):\n  \u003Cverbatim insertions from assets\u002Fsnippets\u002F with source_url>\n\nCOVERAGE CAVEAT:\n  Doctrine dissected on 2026-05-02. 189 cited rules. Gaps in references\u002Fgaps.md.\n  \u003Cif model is INFERRED>: Target model profile inherited; confidence: low.\n```\n\nVerbose adds: per-finding reasoning trail, preservation log (what stayed and\nwhy), and explicit open questions for authorial decisions.\n\n## Workflow\n\n1. Run `python scripts\u002Fclassify_artifact.py \u003Cpath>` → `{type, line_count, signals}`.\n2. Read `references\u002Frubric-by-type.yaml` for the type's rule_ids.\n3. Run `python scripts\u002Fpass1_mechanical.py \u003Cpath> --type \u003Ctype> --target-model \u003Cmodel>`\n   → deterministic findings (including hybrid candidates).\n4. For each rule with `detection_method=llm-judge` in the type's subset,\n   read the relevant section, reason about adherence, cite the rule's\n   `verbatim_quote`, and append the finding to the bundle.\n5. Apply D9 emphasis-conflict resolution by artifact type\n   (see `references\u002Frubric-by-type.yaml` § `false_positive_rules`).\n6. Run `python scripts\u002Fscore_calculator.py \u003Cfindings.json>`\n   → `{score, per_dimension, voice_drift_estimate}`.\n7. Triage findings (🔴\u002F🟡\u002F🟢\u002F❓\u002F⚪).\n8. When mode ≠ `audit`, run `python scripts\u002Foptimize.py \u003Cpath> --findings \u003Cjson> --type \u003Ctype>`\n   to produce surgical diffs and snippet patches (each cited).\n9. When mode = `full`, simulate post-diff, re-score, validate gate.\n10. Emit report (concise default; verbose with `--verbose`).\n\n## Scope discipline (positive framing)\n\nThis skill audits prompts and Claude-directing artifacts. Its responsibilities:\n\n- **Audits**: cited findings only, with `verbatim_quote` + `source_url`.\n- **Proposes diffs**: surgical, one rule per diff, voice-preserving.\n- **Stays in source language**: preserves the artifact's authorial voice.\n- **Defers to operator**: the operator stays in the editor's seat — diffs\n  apply only with `--apply` to an `\u003Cpath>.optimized.\u003Cext>` sibling, never\n  overwriting the original.\n\nOut-of-scope (sibling tools handle these):\n\n- Evaluating Claude's *responses* — that work belongs to eval pipelines.\n- Auditing application code that *consumes* the Claude API beyond the\n  prompts and artifacts themselves.\n- Inventing rules without a source URL — every finding ships cited.\n\n## Edge cases\n\nWhen a rule could fire but no verbatim source quote can be found, hold the\nfinding back. Add it to the report's `coverage_caveat` block as a candidate\nfor the next dissection round. Anthropic-grade means cited or silent.\n\nWhen two rules conflict and no documented resolution exists (see\n`references\u002Fgaps.md` for the 10 known conflicts), surface both findings and\nelevate the decision to the **open questions** block — the operator decides.\n\n## Self-audit: Open Questions\n\nThe skill audits itself on every release via the canonical entry point:\n\n```\npython scripts\u002Frun.py SKILL.md --target opus-4-7 --mode audit\n```\n\nDeclared exceptions (operator-justified, registered in `scripts\u002Frun.py § DECLARED_EXCEPTIONS`):\n\n- **AR-CC-S14 — reserved word \"anthropic\" in name.** The doctrine quote\n  *\"Cannot contain reserved words: 'anthropic', 'claude'.\"* applies to skill\n  names. The operator decision is to keep the name\n  `anthropic-grade-optimizer` because it is the most precise descriptor of\n  the skill's purpose. A rename to `doctrine-grade-optimizer` would lose the\n  explicit doctrinal anchor and degrade trigger reliability for operators\n  searching by domain. The Pass-1 detector continues to flag it on every\n  run; the orchestrator splits the report into `active score` (exception\n  applied) and `raw score` (`--no-exceptions`) so consumers retain full\n  information. D-CC dimension shows the residual penalty\n  transparently — voice fidelity over hidden adjustment.\n\nThe full self-audit run, including findings, both scores, and dimension\nbreakdown, is committed at `evals\u002FSELF-AUDIT.md` and re-runs on every release.\n\n## References (10 files in `references\u002F`)\n\n- `rules-anthropic.yaml` — 189 rules with `source_url` + `verbatim_quote` +\n  `detection_method` (v1.2). The canonical SSOT for rule count and content.\n- `rules-master.yaml` — index by dimension (points into rules-anthropic.yaml).\n- `modulation-matrix.yaml` — 4 models × ~10 cells, each\n  `{value, source_url, confidence}`.\n- `rubric-by-type.yaml` — rule subsets per artifact type, with\n  false-positive suppression matrix and 3 documented doctrinal conflicts.\n- `quotes-canonical.md` — verbatim citations for report output.\n- `heuristics.md` — H1-H12 operational shortcuts.\n- `anti-patterns.md` — AP-1 to AP-15 with detection method.\n- `gaps.md` — known doctrinal gaps (declare in every report).\n- `pass2-protocol.md` — qualitative auditing workflow (llm-judge rules).\n\n## Assets (canonical snippets)\n\n- `assets\u002Fsnippets\u002F` — 18 canonical snippets (16 rule-bound + 2 advisory).\n  Every snippet is byte-by-byte verbatim from the live Anthropic doc, with\n  a header table declaring `triggers_rule`, `applies_to`, `model_specific`,\n  `source_url`, `insertion_pattern`, and `voice_drift`. Snippets cover\n  high-leverage rules: AR-MODEL-018 (code-review coverage), AR-MODEL-011\n  (investigate before answering), AR-MODEL-007\u002FAR-AGENT-009 (parallel\n  tool calls), AR-CLAR-011 (action vs research posture, two snippets),\n  and 12 more.\n- `assets\u002Fsnippets\u002Findex.yaml` — `rule_id → snippet` map consumed by\n  `scripts\u002Foptimize.py`. Type-aware suppression prevents AP-9 false patches\n  (e.g., subagent files never receive subagent-orchestration snippets).\n\n## Scripts (6 entry points in `scripts\u002F`)\n\n- `run.py` — orchestrator; runs classify → pass 1 → score → emit. Single\n  entry point for most workflows.\n- `classify_artifact.py \u003Cpath>` — returns\n  `{type, line_count, language, has_frontmatter, signals}`.\n- `pass1_mechanical.py \u003Cpath> --type \u003Ctype> --target-model \u003Cmodel>` —\n  emits findings array (regex + code-check + hybrid candidates).\n- `score_calculator.py \u003Cfindings.json>` — emits\n  `{score, per_dimension, voice_drift_estimate}`.\n- `optimize.py \u003Cpath> --findings \u003Cjson> --type \u003Ctype> [--apply]` — emits\n  diffs and snippet patches; with `--apply` writes to\n  `\u003Cpath>.optimized.\u003Cext>` (preserving the original).\n- `run_eval_suite.py` — replays `evals\u002Ffixtures\u002F` ground-truth suite.\n\nAll scripts use forward slashes and read paths via the `CLAUDE_SKILL_DIR`\nenv var, defaulting to the parent of the script directory (per AR-CC-S06).\nOn Windows, scripts force UTF-8 stdout to render the triage glyphs\n(🔴\u002F🟡\u002F🟢\u002F❓\u002F⚪) without `cp1252` crashes.\n\n## Validation (eval suite + strange-loop)\n\nThe skill validates itself via `evals\u002F`. The fixture set covers good \u002F\nbad \u002F adversarial-minimal cases for every artifact type, plus cross-type\nfixtures that exercise the false-positive suppression matrix. Ground-truth\nscores live in `evals\u002Fground-truth.yaml`; `scripts\u002Frun_eval_suite.py`\nreplays them. The strange-loop self-audit\n(`evals\u002FSELF-AUDIT.md`) re-runs the skill against its own SKILL.md on every\nrelease and commits the result for transparency.\n",{"name":5,"description":7,"when_to_use":61},"Operator asks to audit, optimize, review, lint, score, refactor, or \"make Anthropic-grade\" any prompt, CLAUDE.md, SKILL.md, subagent, hook, MCP config, api_config, or workflow. Also triggers on model migration (\"good for Opus 4.7?\", \"optimize for Haiku\"), compliance review, or fixing adherence\u002Fverbosity\u002Fhallucination drift in a Claude artifact.","main","63188abc441e3da9c6541a6d21f548a0d446a3ec","",8,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer","https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1","Building AI-native Cognitive Systems.\r\n| Founder @ AIdeas Lab.\r\n| Architect of latent cognition.","2026-11-06 04:47:21",{"source":8,"skillName":5,"commands":71,"prompt":81,"zipAvailable":82},[72,75,78],{"tool":73,"command":74},"npx","npx skills add l0z4n0-a1\u002Fskill-anthropic-grade-optimizer --skill anthropic-grade-optimizer",{"tool":76,"command":77},"bunx","bunx skills add l0z4n0-a1\u002Fskill-anthropic-grade-optimizer --skill anthropic-grade-optimizer",{"tool":79,"command":80},"pnpm","pnpm dlx skills add l0z4n0-a1\u002Fskill-anthropic-grade-optimizer --skill anthropic-grade-optimizer","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FSKILL.md",true,[84,90,96,102,106,110,114,118,122,126,130,134,138,142,146,150,154,158,162,166,170,174,178,182,186,190,194,198,202,206,210,214,218,222,226,230,234,238,242,246,250,254,258,262,266,270,274,278,282,286,290,294,298,302,306,310,316,320,324,328,332,336,340,344,348,352,356,360,364,368,372],{"path":85,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":88,"sourceHtmlUrl":89},".github\u002Fworkflows\u002Fci.yml","REFERENCE","参考文档",1239,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002F.github\u002Fworkflows\u002Fci.yml",{"path":91,"resourceType":92,"resourceTypeDesc":93,"sizeBytes":94,"sourceHtmlUrl":95},".gitignore","OTHER","其他",664,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002F.gitignore",{"path":97,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":100,"sourceHtmlUrl":101},"assets\u002Fsnippets\u002Favoid-excessive-markdown.md","ASSET","资产文件",1995,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Favoid-excessive-markdown.md",{"path":103,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":104,"sourceHtmlUrl":105},"assets\u002Fsnippets\u002Favoid-overengineering.md",1876,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Favoid-overengineering.md",{"path":107,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":108,"sourceHtmlUrl":109},"assets\u002Fsnippets\u002Fcleanup-temp-files.md",1188,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fcleanup-temp-files.md",{"path":111,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":112,"sourceHtmlUrl":113},"assets\u002Fsnippets\u002Fcode-review-coverage.md",2304,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fcode-review-coverage.md",{"path":115,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":116,"sourceHtmlUrl":117},"assets\u002Fsnippets\u002Fconcise-output.md",1275,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fconcise-output.md",{"path":119,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":120,"sourceHtmlUrl":121},"assets\u002Fsnippets\u002Fdefault-to-action.md",1197,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fdefault-to-action.md",{"path":123,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":124,"sourceHtmlUrl":125},"assets\u002Fsnippets\u002Fdo-not-act-before-instructions.md",1200,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fdo-not-act-before-instructions.md",{"path":127,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":128,"sourceHtmlUrl":129},"assets\u002Fsnippets\u002Ffrontend-aesthetics.md",1503,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Ffrontend-aesthetics.md",{"path":131,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":132,"sourceHtmlUrl":133},"assets\u002Fsnippets\u002Fgeneral-purpose-solution.md",1783,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fgeneral-purpose-solution.md",{"path":135,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":136,"sourceHtmlUrl":137},"assets\u002Fsnippets\u002Findex.yaml",4827,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Findex.yaml",{"path":139,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":140,"sourceHtmlUrl":141},"assets\u002Fsnippets\u002Finvestigate-before-answering.md",1441,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Finvestigate-before-answering.md",{"path":143,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":144,"sourceHtmlUrl":145},"assets\u002Fsnippets\u002Flatex-opt-out.md",1221,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Flatex-opt-out.md",{"path":147,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":148,"sourceHtmlUrl":149},"assets\u002Fsnippets\u002Fmodel-self-knowledge.md",1112,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fmodel-self-knowledge.md",{"path":151,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":152,"sourceHtmlUrl":153},"assets\u002Fsnippets\u002Fmulti-window-context.md",1526,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fmulti-window-context.md",{"path":155,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":156,"sourceHtmlUrl":157},"assets\u002Fsnippets\u002Fpick-and-commit.md",1349,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fpick-and-commit.md",{"path":159,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":160,"sourceHtmlUrl":161},"assets\u002Fsnippets\u002Freversibility-safety.md",1607,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Freversibility-safety.md",{"path":163,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":164,"sourceHtmlUrl":165},"assets\u002Fsnippets\u002Fsteer-thinking.md",1316,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fsteer-thinking.md",{"path":167,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":168,"sourceHtmlUrl":169},"assets\u002Fsnippets\u002Fsubagent-spawning-4-7.md",1332,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fsubagent-spawning-4-7.md",{"path":171,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":172,"sourceHtmlUrl":173},"assets\u002Fsnippets\u002Fuse-parallel-tool-calls.md",1571,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fassets\u002Fsnippets\u002Fuse-parallel-tool-calls.md",{"path":175,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":176,"sourceHtmlUrl":177},"CHANGELOG.md",3758,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FCHANGELOG.md",{"path":179,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":180,"sourceHtmlUrl":181},"CONTRIBUTING.md",3722,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FCONTRIBUTING.md",{"path":183,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":184,"sourceHtmlUrl":185},"evals\u002Ffixtures\u002Fadversarial-minimal-claude-md.md",76,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fadversarial-minimal-claude-md.md",{"path":187,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":188,"sourceHtmlUrl":189},"evals\u002Ffixtures\u002Fadversarial-minimal-skill.md",276,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fadversarial-minimal-skill.md",{"path":191,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":192,"sourceHtmlUrl":193},"evals\u002Ffixtures\u002Fadversarial-minimal-subagent.md",241,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fadversarial-minimal-subagent.md",{"path":195,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":196,"sourceHtmlUrl":197},"evals\u002Ffixtures\u002Fbad-claude-md.md",3679,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fbad-claude-md.md",{"path":199,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":200,"sourceHtmlUrl":201},"evals\u002Ffixtures\u002Fbad-hook.json",492,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fbad-hook.json",{"path":203,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":204,"sourceHtmlUrl":205},"evals\u002Ffixtures\u002Fbad-mcp.json",279,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fbad-mcp.json",{"path":207,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":208,"sourceHtmlUrl":209},"evals\u002Ffixtures\u002Fbad-skill.md",558,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fbad-skill.md",{"path":211,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":212,"sourceHtmlUrl":213},"evals\u002Ffixtures\u002Fbad-subagent.md",402,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fbad-subagent.md",{"path":215,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":216,"sourceHtmlUrl":217},"evals\u002Ffixtures\u002Fcross-bad-skill.md",620,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fcross-bad-skill.md",{"path":219,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":220,"sourceHtmlUrl":221},"evals\u002Ffixtures\u002Fcross-bad-subagent.md",216,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fcross-bad-subagent.md",{"path":223,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":224,"sourceHtmlUrl":225},"evals\u002Ffixtures\u002Fgood-claude-md.md",851,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fgood-claude-md.md",{"path":227,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":228,"sourceHtmlUrl":229},"evals\u002Ffixtures\u002Fgood-hook.json",538,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fgood-hook.json",{"path":231,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":232,"sourceHtmlUrl":233},"evals\u002Ffixtures\u002Fgood-mcp.json",368,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fgood-mcp.json",{"path":235,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":236,"sourceHtmlUrl":237},"evals\u002Ffixtures\u002Fgood-skill.md",1449,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fgood-skill.md",{"path":239,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":240,"sourceHtmlUrl":241},"evals\u002Ffixtures\u002Fgood-subagent.md",1248,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Ffixtures\u002Fgood-subagent.md",{"path":243,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":244,"sourceHtmlUrl":245},"evals\u002Fground-truth.yaml",4251,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Fground-truth.yaml",{"path":247,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":248,"sourceHtmlUrl":249},"evals\u002FREADME.md",1909,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002FREADME.md",{"path":251,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":252,"sourceHtmlUrl":253},"evals\u002Fresults\u002FCALIBRATION_REPORT.md",2242,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Fresults\u002FCALIBRATION_REPORT.md",{"path":255,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":256,"sourceHtmlUrl":257},"evals\u002Fresults\u002FEVAL_MATRIX.md",1242,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002Fresults\u002FEVAL_MATRIX.md",{"path":259,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":260,"sourceHtmlUrl":261},"evals\u002FSELF-AUDIT.md",7874,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002FSELF-AUDIT.md",{"path":263,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":264,"sourceHtmlUrl":265},"evals\u002FSTRANGE-LOOP-RESULTS.md",5397,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fevals\u002FSTRANGE-LOOP-RESULTS.md",{"path":267,"resourceType":267,"resourceTypeDesc":268,"sizeBytes":269,"sourceHtmlUrl":16},"LICENSE","许可证",1077,{"path":271,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":272,"sourceHtmlUrl":273},"README.md",8693,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FREADME.md",{"path":275,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":276,"sourceHtmlUrl":277},"references\u002Fanti-patterns.md",6323,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Fanti-patterns.md",{"path":279,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":280,"sourceHtmlUrl":281},"references\u002Fgaps.md",7945,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Fgaps.md",{"path":283,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":284,"sourceHtmlUrl":285},"references\u002Fheuristics.md",4002,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Fheuristics.md",{"path":287,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":288,"sourceHtmlUrl":289},"references\u002Fmodulation-matrix.yaml",6911,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Fmodulation-matrix.yaml",{"path":291,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":292,"sourceHtmlUrl":293},"references\u002Fpass2-protocol.md",7607,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Fpass2-protocol.md",{"path":295,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":296,"sourceHtmlUrl":297},"references\u002Fquotes-canonical.md",6390,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Fquotes-canonical.md",{"path":299,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":300,"sourceHtmlUrl":301},"references\u002Frubric-by-type.yaml",22099,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Frubric-by-type.yaml",{"path":303,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":304,"sourceHtmlUrl":305},"references\u002Frules-anthropic.yaml",108753,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Frules-anthropic.yaml",{"path":307,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":308,"sourceHtmlUrl":309},"references\u002Frules-master.yaml",4897,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Freferences\u002Frules-master.yaml",{"path":311,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":314,"sourceHtmlUrl":315},"scripts\u002Fcheck_model_currency.py","SCRIPT","脚本",4253,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Fcheck_model_currency.py",{"path":317,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":318,"sourceHtmlUrl":319},"scripts\u002Fclassify_artifact.py",5434,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Fclassify_artifact.py",{"path":321,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":322,"sourceHtmlUrl":323},"scripts\u002Flib_ux.py",18354,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Flib_ux.py",{"path":325,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":326,"sourceHtmlUrl":327},"scripts\u002Foptimize.py",22525,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Foptimize.py",{"path":329,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":330,"sourceHtmlUrl":331},"scripts\u002Fpass1_mechanical.py",69290,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Fpass1_mechanical.py",{"path":333,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":334,"sourceHtmlUrl":335},"scripts\u002Fpass2_executor.py",10714,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Fpass2_executor.py",{"path":337,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":338,"sourceHtmlUrl":339},"scripts\u002Frun_eval_suite.py",7952,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Frun_eval_suite.py",{"path":341,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":342,"sourceHtmlUrl":343},"scripts\u002Frun.py",17755,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Frun.py",{"path":345,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":346,"sourceHtmlUrl":347},"scripts\u002Fscore_calculator.py",3932,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Fscripts\u002Fscore_calculator.py",{"path":349,"resourceType":86,"resourceTypeDesc":87,"sizeBytes":350,"sourceHtmlUrl":351},"SECURITY.md",1543,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FSECURITY.md",{"path":353,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":354,"sourceHtmlUrl":355},"tests\u002F__init__.py",0,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Ftests\u002F__init__.py",{"path":357,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":358,"sourceHtmlUrl":359},"tests\u002Fcalibration.py",8588,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Ftests\u002Fcalibration.py",{"path":361,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":362,"sourceHtmlUrl":363},"tests\u002Ftest_orchestrator.py",3589,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Ftests\u002Ftest_orchestrator.py",{"path":365,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":366,"sourceHtmlUrl":367},"tests\u002Ftest_pass1_detectors.py",3414,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Ftests\u002Ftest_pass1_detectors.py",{"path":369,"resourceType":312,"resourceTypeDesc":313,"sizeBytes":370,"sourceHtmlUrl":371},"tests\u002Ftest_rule_schema.py",3303,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002Ftests\u002Ftest_rule_schema.py",{"path":373,"resourceType":92,"resourceTypeDesc":93,"sizeBytes":374,"sourceHtmlUrl":375},"VERSION",6,"https:\u002F\u002Fgithub.com\u002Fl0z4n0-a1\u002Fskill-anthropic-grade-optimizer\u002Fblob\u002Fmain\u002FVERSION",[]]