[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6KD9CK9V2JBA6B8EBK8CWXC":3,"skill-same:01M6KD9CK9V2JBA6B8EBK8CWXC":68},{"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":6,"sourceHtmlUrl":15,"categories":16,"keywords":22,"publishedAt":43,"linkOnly":44,"markdownContent":45,"markdownBody":46,"frontMatter":47,"refName":48,"commitSha":49,"skillDirectory":50,"forksCount":13,"repositoryHtmlUrl":51,"repositoryHomepageUrl":6,"ownerHtmlUrl":52,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":53,"install":54,"resources":67},"01M6KD9CK9V2JBA6B8EBK8CWXC","system-prompts",null,"Write system prompts, tool docs, and agent definitions. Combines research-backed prompt engineering (+15-30% measured improvements) with project XML conventions. Covers tag hierarchy, structural templates, high-impact interventions, anti-patterns.","mykie2015\u002Foh-my-pi-node","mykie2015","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F9046885?v=4","01M6KD754M17CQWKRW1ZK3X2XG","01M6KD758Y55WPRWDTTYZV84Z6",0,"MIT","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002F.omp\u002Fskills\u002Fsystem-prompts\u002FSKILL.md",[17],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42],"system prompts","system","prompts","hierarchy","docs","high","with","impact","structural","project","conventions","patterns","anti-patterns","write","prompt","tool","templates","xml","research","prompt engineering","2026-11-10 10:12:30",false,"---\nname: system-prompts\ndescription: Write system prompts, tool docs, and agent definitions. Combines research-backed prompt engineering (+15-30% measured improvements) with project XML conventions. Covers tag hierarchy, structural templates, high-impact interventions, anti-patterns.\n---\n\n# System Prompt Engineering\n\nEmpirically-validated techniques + consistent XML structure. Every recommendation backed by benchmarks or production data.\n\n\u003Ccritical>\n## High-Impact Interventions (+15-30% measured improvement)\n\n1. **Persistence**: \"Keep going until fully resolved\" — prevents premature termination\n2. **Tool verification**: \"Use tools to verify; do not guess\" — reduces hallucination\n3. **Planning**: \"Plan approach before acting\" — improves complex task success\n4. **Context positioning**: Critical instructions at START and END — middle content degrades 20%+\n5. **Urgency framing**: \"This matters\" \u002F \"Get this right\" — 8-115% improvement (EmotionPrompt)\n6. **Edit format**: SEARCH\u002FREPLACE beats line-numbers 3X on code generation\n\n**Minimal prompting wins.** Every instruction must justify its token cost.\n\u003C\u002Fcritical>\n\n---\n\n## Tag Hierarchy\n\nTags encode enforcement level. Use consistently throughout:\n\n| Tag             | Enforcement   | When to Use                                          |\n| --------------- | ------------- | ---------------------------------------------------- |\n| `\u003Ccritical>`    | Inviolable    | Safety constraints, must-follow rules, repeat at END |\n| `\u003Cprohibited>`  | Forbidden     | Actions that cause harm, never acceptable            |\n| `\u003Ccaution>`     | High priority | Important to follow                                  |\n| `\u003Cinstruction>` | Operational   | How to use a tool, perform a task                    |\n| `\u003Cconditions>`  | Contextual    | When rules apply, trigger criteria                   |\n| `\u003Cavoid>`       | Anti-patterns | What not to do, prefer alternatives                  |\n\n**Context positioning rule**: Place `\u003Ccritical>` at START for immediate priming, repeat at END for recency. Middle content suffers 20%+ degradation in long contexts.\n\n---\n\n## Standard Tags\n\n### Structure Tags\n\n```\n\u003Crole>           Agent identity and expertise (first element)\n\u003Ccontext>        Background, situation, audience\n\u003Cprocedure>      Numbered step-by-step workflows\n\u003Cdirectives>     Bulleted operating instructions\n\u003Cparameters>     Input specifications, types\n\u003Coutput>         Return value documentation\n\u003Cstrengths>      What the agent\u002Ftool excels at\n\u003Coperations>     Available operations (for multi-op tools like LSP)\n```\n\n### Special Tags\n\n```\n\u003Cnorth-star>     Core values, ultimate objectives\n\u003Cstance>         Communication style, attitude\n\u003Ccommitment>     What the agent commits to doing\n\u003Cfield>          Domain-specific mindset\u002Fcontext\n\u003Cprotocol>       Behavioral rules, tool precedence\n```\n\n### Example Tags\n\nAlways use `name` attribute with lowercase-kebab descriptive names:\n\n```xml\n\u003Cexample name=\"good\">\nClear, correct usage\n\u003C\u002Fexample>\n\n\u003Cexample name=\"bad\">\nWhat to avoid — show the mistake explicitly\n\u003C\u002Fexample>\n\n\u003Cexample name=\"rate-limiting\">\nDomain-specific example\n\u003C\u002Fexample>\n\n\u003Cexample name=\"windows-cmd\">\nPlatform\u002Fcontext-specific\n\u003C\u002Fexample>\n```\n\nNaming patterns:\n\n- `name=\"single\"` \u002F `name=\"multi-part\"` — complexity variants\n- `name=\"good\"` \u002F `name=\"bad\"` — correctness contrast\n- `name=\"linux\"` \u002F `name=\"windows-cmd\"` — platform-specific\n- `name=\"create\"` \u002F `name=\"update\"` \u002F `name=\"delete\"` — operation types\n\n---\n\n## Structural Templates\n\n### Tool Documentation\n\n```markdown\n# Tool Name\n\nOne-line description of what the tool does.\n\n\u003Cinstruction>\n- How to use it (bulleted, imperative)\n- Key parameters and their effects\n- Common patterns\n\u003C\u002Finstruction>\n\n\u003Coutput>\nWhat the tool returns. Include:\n- Success format\n- Truncation limits (e.g., \"truncated at 50KB\")\n- Error conditions\n\u003C\u002Foutput>\n\n\u003Ccritical>\nMust-follow rules. Safety constraints.\nWhen to ALWAYS or NEVER use this tool.\n\u003C\u002Fcritical>\n\n\u003Ccaution>\nHigh-priority notes that aren't safety-critical.\n\u003C\u002Fcaution>\n\n\u003Cexample name=\"basic\">\ntool {\"param\": \"value\"}\n\u003C\u002Fexample>\n\n\u003Cexample name=\"advanced\">\ntool {\"param\": \"value\", \"option\": true}\n\u003C\u002Fexample>\n\n\u003Cavoid>\n- Anti-pattern 1 — why it's bad\n- Anti-pattern 2 — what to do instead\n\u003C\u002Favoid>\n```\n\n### Agent Definition\n\n```markdown\n---\nname: agent-name\ndescription: One-line for spawning UI (imperative: \"Fast read-only codebase scout\")\ntools: read, grep, find, bash\nmodel: pi\u002Fslow, gpt-5.2, codex\noutput:\n  properties:\n    field_name:\n      metadata:\n        description: What this field contains\n      type: string\n---\n\n\u003Crole>Senior [role] doing [task]. Your goal: [concrete outcome].\u003C\u002Frole>\n\n\u003Ccritical>\nInviolable constraints first.\nREAD-ONLY if applicable — list prohibited actions explicitly.\n\u003C\u002Fcritical>\n\n\u003Cstrengths>\n- What this agent excels at\n- Core capabilities\n\u003C\u002Fstrengths>\n\n\u003Cdirectives>\n- Operating instruction 1\n- Operating instruction 2\n- Spawn parallel tool calls wherever possible\n\u003C\u002Fdirectives>\n\n\u003Cprocedure>\n## Phase 1: Understand\n1. Step one\n2. Step two\n\n## Phase 2: Execute\n\n1. Step one\n2. Step two\n   \u003C\u002Fprocedure>\n\n\u003Coutput>\nWhat to return. Schema requirements.\nCall `submit_result` with findings when done.\n\u003C\u002Foutput>\n\n\u003Ccritical>\nRepeat critical constraints at end.\nKeep going until complete. This matters.\n\u003C\u002Fcritical>\n```\n\n### System Prompt (Main Agent)\n\n```markdown\n\u003Csystem-directive>\nXML tags in this prompt are system-level instructions. They are not suggestions.\n\nTag hierarchy (by enforcement level):\n\n- `\u003Ccritical>` — Inviolable. Failure to comply is a system failure.\n- `\u003Cprohibited>` — Forbidden. These actions will cause harm.\n- `\u003Ccaution>` — High priority. Important to follow.\n- `\u003Cinstruction>` — How to operate. Follow precisely.\n- `\u003Cconditions>` — When rules apply. Check before acting.\n- `\u003Cavoid>` — Anti-patterns. Prefer alternatives.\n  \u003C\u002Fsystem-directive>\n\nYou are a [specific role with credentials].\n\n\u003Cfield>\nDomain-specific context and mindset.\nWhat to notice, what traps exist.\n\u003C\u002Ffield>\n\n\u003Cstance>\nCommunication style.\nCorrectness over politeness. Brevity over ceremony.\n\u003C\u002Fstance>\n\n\u003Cprotocol>\n## Tool Precedence\nSpecialized tools → Python → Bash\n...\n\n## Verification\n\nExternal proof: tests, linters, type checks.\n...\n\u003C\u002Fprotocol>\n\n\u003Cprocedure>\n## Before action\n1. CHECKPOINT — pause, assess parallelism\n2. Plan if task has weight\n3. State intent before each tool call\n\u003C\u002Fprocedure>\n\n\u003Cnorth-star>\nCore values. What ultimately matters.\n\u003C\u002Fnorth-star>\n\n\u003Cprohibited>\nActions that cause harm.\n\u003C\u002Fprohibited>\n\n\u003Ccritical>\nRepeat most important rules.\nKeep going until finished.\nThe work is done when it is correct.\n\u003C\u002Fcritical>\n```\n\n---\n\n## Writing Style\n\n### Voice\n\n**Direct and imperative.** Research shows direct tone improves accuracy 4%+ over polite hedging.\n\n```\nBad:  \"You might want to consider using...\"\nGood: \"Use X when Y.\"\n\nBad:  \"It would be helpful if you could...\"\nGood: \"Do X.\"\n\nBad:  \"Please note that this is important...\"\nGood: \"Critical: X.\"\n```\n\n**Urgency framing** (8-115% improvement):\n\n```\n\"This matters. Get it right.\"\n\"Be thorough.\"\n\"Keep going until fully resolved.\"\n```\n\n### Normative Language (RFC 2119)\n\nAll prompt prose that prescribes behavior MUST use RFC 2119 key words in **full caps**. This removes ambiguity about whether an instruction is absolute or advisory.\n\n| Keyword | Meaning | Replaces |\n| --- | --- | --- |\n| **MUST** \u002F **REQUIRED** | Absolute requirement | \"always\", \"make sure\", \"ensure\", \"do\" |\n| **MUST NOT** \u002F **PROHIBITED** | Absolute prohibition | \"never\", \"do not\", \"don't\", \"strictly prohibited\" |\n| **SHOULD** \u002F **RECOMMENDED** | Strong preference; deviation allowed with known tradeoffs | \"prefer\", \"recommend\", \"it's best to\" |\n| **SHOULD NOT** \u002F **NOT RECOMMENDED** | Strong discouragement; deviation allowed with known tradeoffs | \"avoid\", \"try not to\" |\n| **MAY** \u002F **OPTIONAL** | Truly optional | \"can\", \"may\", \"you could\" |\n\n```\nBad:  \"Never edit from a grep snippet alone\"\nGood: \"You MUST NOT edit from a grep snippet alone\"\n\nBad:  \"Prefer unit tests over mocks\"\nGood: \"You SHOULD prefer unit tests over mocks\"\n\nBad:  \"Make sure to run lsp references before modifying a symbol\"\nGood: \"You MUST run lsp references before modifying any symbol\"\n```\n\n**What not to convert**: factual\u002Fdescriptive sentences (what a tool returns, what a parameter does), code blocks, examples, schema definitions, Handlebars template syntax. Only prescriptive prose gets RFC treatment.\n\n### Positive Framing\n\nModels process \"Always do Y\" better than \"Don't do X\":\n\n```\nBad:  \"Don't use grep via bash\"\nGood: \"ALWAYS use Grep tool for search—NEVER invoke grep via Bash\"\n\nBad:  \"Don't guess\"\nGood: \"Use tools to verify; do not guess\"\n```\n\nWhen negation is necessary, pair with positive alternative.\n\n### Specificity\n\n**Role specificity spectrum** (effectiveness increases →):\n\n```\n\"You are a lawyer\"\n    ↓\n\"You are a corporate M&A lawyer\"\n    ↓\n\"You are General Counsel at a Fortune 500 tech company, 15 years in SaaS licensing\"\n```\n\n**Constraint specificity**:\n\n```\nBad:  \"Keep it short\"\nGood: \"3 bullets, \u003C50 words each\"\n\nBad:  \"Be careful with large files\"\nGood: \"Truncated at 50KB or 2000 lines, whichever comes first\"\n```\n\n### Formatting\n\n```\n# H1 for tool\u002Fagent name only\n## H2 for major sections\n### H3 sparingly\n\n- Bullets for unordered lists inside tags\n1. Numbers for ordered procedures\n\n| Tables | For | Structured reference data |\n\n`inline code` for commands, paths, parameters, values\n```\n\nCode blocks with language:\n\n````markdown\n```typescript\nconst example = \"always specify language\";\n```\n````\n\n---\n\n## Technique Reference\n\n### Chain of Thought\n\n**Use when**: Multi-step reasoning, math, analysis, complex decisions\n**Avoid when**: Simple tasks, reasoning models (o1\u002Fo3 do internal CoT)\n\n```xml\n\u003Cinstruction>\nBefore answering:\n1. Identify the core question\n2. List relevant constraints\n3. Consider 2-3 approaches\n4. Select best with rationale\n\nThen provide your answer.\n\u003C\u002Finstruction>\n```\n\n**Token-efficient variant** (Chain of Draft):\n\n```\nThink step-by-step, keeping only 5-word notes per step.\nOutput final answer after ####.\n```\n\n### Few-Shot Examples\n\n**Use when**: Enforcing specific output format, classification, smaller models\n**Avoid when**: Advanced models (Claude 3.5+, GPT-4+) on clear tasks — adds noise\n\nWhen using: 3-5 diverse examples covering edge cases.\n\n```xml\n\u003Cexample name=\"simple\">\nInput: X\nOutput: Y\n\u003C\u002Fexample>\n\n\u003Cexample name=\"edge-case\">\nInput: X'\nOutput: Y'\n\u003C\u002Fexample>\n```\n\n### Long Context Handling\n\n**\"Lost in the Middle\"**: Beginning and end retain; middle degrades 20%+.\n\n1. Documents at TOP, instructions AFTER\n2. Quote grounding: \"Quote relevant passages in `\u003Cquotes>`, then analyze\"\n3. Critical instructions at START and END\n4. Chunk >100K tokens → process parallel → synthesize\n\n```xml\n\u003Cdocuments>\n{{LONG_CONTENT}}\n\u003C\u002Fdocuments>\n\n\u003Cinstructions>\n1. Find and quote passages relevant to {{QUERY}} in \u003Cquotes>\n2. Analyze based on quoted evidence in \u003Canalysis>\n\u003C\u002Finstructions>\n```\n\n### Verification Patterns\n\n**Self-correction without external feedback does not work.**\n\nEffective:\n\n```\n1. Generate solution\n2. Execute verification (tests, lint, typecheck)\n3. On failure: analyze error → fix → re-verify\n4. Iterate until pass\n```\n\nIneffective:\n\n```\n1. Generate solution\n2. \"Critique your solution\"      ← detection is the bottleneck\n3. \"Improve based on critique\"   ← feels productive, doesn't help\n```\n\n### Prompt Chaining\n\n**Use when**: Single prompt drops steps, distinct phases, verification needed between.\n\n```\nPrompt 1: Analyze → \u003Canalysis>\nPrompt 2: \u003Canalysis> → Plan → \u003Cplan>\nPrompt 3: \u003Cplan> → Execute → \u003Cresult>\nPrompt 4: \u003Cresult> → Verify → final\n```\n\n---\n\n## Anti-Patterns (Measured Degradation)\n\n| Pattern                                   | Problem                                 |\n| ----------------------------------------- | --------------------------------------- |\n| \"Would you be so kind...\"                 | +perplexity, -4% accuracy               |\n| \"I'll tip $2000\"                          | No improvement, sometimes worse         |\n| Explicit CoT on reasoning models (o1\u002Fo3)  | -36%, conflicts with internal reasoning |\n| Few-shot on advanced models + clear tasks | Introduces noise\u002Fbias                   |\n| \"Always end with Progress\u002FQuestions\"      | Degrades task performance               |\n| \"Be efficient with tokens\"                | Premature task abandonment              |\n| \"Don't do X\" without positive alternative | \"Always do Y\" processes better          |\n| Verbose explanations of obvious concepts  | Context bloat; model already knows      |\n| Self-critique without external feedback   | Detection is bottleneck, not correction |\n| Critical instructions only in middle      | 20%+ degradation vs start\u002Fend           |\n\n---\n\n## Complete Examples\n\n### Tool Doc Example\n\n```markdown\n# Grep\n\nFast regex search built on ripgrep.\n\n\u003Cinstruction>\n- Full regex: `log.*Error`, `function\\\\s+\\\\w+`\n- Filter: `glob` (e.g., `*.js`) or `type` (e.g., `js`, `py`)\n- Cross-line: `multiline: true` for patterns like `struct \\\\{[\\\\s\\\\S]*?field`\n\u003C\u002Finstruction>\n\n\u003Coutput>\nDepends on `output_mode`:\n- `content`: Lines with paths and line numbers (default limit: 100)\n- `files_with_matches`: Paths only\n- `count`: Match counts per file\n\nTruncated results reference `artifact:\u002F\u002F\u003Cid>` for full output.\n\u003C\u002Foutput>\n\n\u003Ccritical>\nALWAYS use Grep for search—NEVER invoke `grep` or `rg` via Bash.\n\u003C\u002Fcritical>\n\n\u003Cexample name=\"regex\">\ngrep {\"pattern\": \"function\\\\s+\\\\w+\", \"glob\": \"*.ts\"}\n\u003C\u002Fexample>\n\n\u003Cexample name=\"multiline\">\ngrep {\"pattern\": \"struct \\\\{[\\\\s\\\\S]*?field\", \"multiline\": true}\n\u003C\u002Fexample>\n\n\u003Cavoid>\n- Open-ended searches requiring multiple rounds—use Task tool\n- Raw bash grep\u002Frg invocation\n\u003C\u002Favoid>\n```\n\n### Agent Example\n\n```markdown\n---\nname: explore\ndescription: Fast read-only codebase scout returning compressed context for handoff\ntools: read, grep, find, bash\nmodel: pi\u002Fsmol, haiku-4.5, haiku-4-5, gemini-flash-latest, gemini-3-flash, zai-glm-4.7, glm-4.7-flash, glm-4.5-flash, gpt-5.1-codex-mini, haiku, flash, mini\noutput:\n  properties:\n    query:\n      type: string\n    files:\n      elements:\n        properties:\n          path: { type: string }\n          line_start: { type: number }\n          line_end: { type: number }\n          description: { type: string }\n    architecture:\n      type: string\n---\n\n\u003Crole>File search specialist. Investigate codebase, return structured findings for handoff.\u003C\u002Frole>\n\n\u003Ccritical>\nREAD-ONLY. You are STRICTLY PROHIBITED from:\n- Creating, editing, deleting files\n- Using redirect operators (>, >>)\n- Running state-changing commands (git add, npm install)\n\u003C\u002Fcritical>\n\n\u003Cstrengths>\n- Rapid file discovery via find patterns\n- Regex search with grep\n- Tracing imports and dependencies\n\u003C\u002Fstrengths>\n\n\u003Cdirectives>\n- Spawn parallel tool calls wherever possible\n- Return absolute paths\n- Communicate findings directly—do NOT create files\n\u003C\u002Fdirectives>\n\n\u003Cprocedure>\n1. grep\u002Ffind to locate relevant code\n2. Read key sections (not entire files)\n3. Identify types, interfaces, key functions\n4. Note dependencies between files\n5. Call `submit_result` with findings\n\u003C\u002Fprocedure>\n\n\u003Ccritical>\nRead-only. Call `submit_result` when done. This matters.\n\u003C\u002Fcritical>\n```\n\n---\n\n\u003Ccritical>\n## Deployment Checklist\n\n- [ ] **Tag hierarchy**: Enforcement level matches content?\n- [ ] **Critical at edges**: Most important rules at START and END?\n- [ ] **Named examples**: All `\u003Cexample>` tags have `name` attribute?\n- [ ] **Positive framing**: \"Do Y\" not just \"Don't X\"?\n- [ ] **Direct tone**: No hedging, no filler, urgency where appropriate?\n- [ ] **Specificity**: Exact formats, limits, constraints—not vague?\n- [ ] **Token efficiency**: Each sentence justifies its cost?\n- [ ] **Verification**: External feedback loop if correctness matters?\n- [ ] **RFC 2119 normative language**: All prescriptive sentences use MUST\u002FMUST NOT\u002FSHOULD\u002FMAY in caps?\n- [ ] **Persistence**: \"Keep going until complete\" for complex tasks?\n\n**High-impact interventions: persistence, tool verification, planning, context positioning, urgency.**\n\u003C\u002Fcritical>\n\n---\n\n## Quick Reference\n\n### Tag Names\n\n```\nEnforcement:  \u003Ccritical> \u003Cprohibited> \u003Ccaution> \u003Cinstruction> \u003Cconditions> \u003Cavoid>\nStructure:    \u003Crole> \u003Ccontext> \u003Cprocedure> \u003Cdirectives> \u003Cparameters> \u003Coutput>\nCapability:   \u003Cstrengths> \u003Ctools> \u003Coperations>\nExamples:     \u003Cexample name=\"kebab-case-name\">\nData:         \u003Cenvironment> \u003Cdata> \u003Cdocuments>\nSpecial:      \u003Cnorth-star> \u003Cstance> \u003Ccommitment> \u003Cfield> \u003Cprotocol>\n```\n\n### Example Name Patterns\n\n```\nCorrectness:  name=\"good\", name=\"bad\"\nComplexity:   name=\"single\", name=\"multi-part\", name=\"basic\", name=\"advanced\"\nOperations:   name=\"create\", name=\"update\", name=\"delete\", name=\"rename\"\nPlatforms:    name=\"linux\", name=\"windows-cmd\", name=\"macos\"\nDomains:      name=\"rate-limiting\", name=\"auth\", name=\"validation\"\n```\n\n### Task → Technique\n\n| Task                | Primary                         | Secondary            |\n| ------------------- | ------------------------------- | -------------------- |\n| Simple extraction   | Clear constraints               | Prefilling           |\n| Classification      | 3-5 examples                    | XML structure        |\n| Complex analysis    | Structured reasoning            | Role + urgency       |\n| Code generation     | SEARCH\u002FREPLACE                  | Verification loop    |\n| Long document       | Docs at top, quote-then-analyze | XML structure        |\n| Multi-step workflow | Prompt chaining                 | Planning instruction |\n| Domain expertise    | Specific role + credentials     | Examples             |\n","# System Prompt Engineering\n\nEmpirically-validated techniques + consistent XML structure. Every recommendation backed by benchmarks or production data.\n\n\u003Ccritical>\n## High-Impact Interventions (+15-30% measured improvement)\n\n1. **Persistence**: \"Keep going until fully resolved\" — prevents premature termination\n2. **Tool verification**: \"Use tools to verify; do not guess\" — reduces hallucination\n3. **Planning**: \"Plan approach before acting\" — improves complex task success\n4. **Context positioning**: Critical instructions at START and END — middle content degrades 20%+\n5. **Urgency framing**: \"This matters\" \u002F \"Get this right\" — 8-115% improvement (EmotionPrompt)\n6. **Edit format**: SEARCH\u002FREPLACE beats line-numbers 3X on code generation\n\n**Minimal prompting wins.** Every instruction must justify its token cost.\n\u003C\u002Fcritical>\n\n---\n\n## Tag Hierarchy\n\nTags encode enforcement level. Use consistently throughout:\n\n| Tag             | Enforcement   | When to Use                                          |\n| --------------- | ------------- | ---------------------------------------------------- |\n| `\u003Ccritical>`    | Inviolable    | Safety constraints, must-follow rules, repeat at END |\n| `\u003Cprohibited>`  | Forbidden     | Actions that cause harm, never acceptable            |\n| `\u003Ccaution>`     | High priority | Important to follow                                  |\n| `\u003Cinstruction>` | Operational   | How to use a tool, perform a task                    |\n| `\u003Cconditions>`  | Contextual    | When rules apply, trigger criteria                   |\n| `\u003Cavoid>`       | Anti-patterns | What not to do, prefer alternatives                  |\n\n**Context positioning rule**: Place `\u003Ccritical>` at START for immediate priming, repeat at END for recency. Middle content suffers 20%+ degradation in long contexts.\n\n---\n\n## Standard Tags\n\n### Structure Tags\n\n```\n\u003Crole>           Agent identity and expertise (first element)\n\u003Ccontext>        Background, situation, audience\n\u003Cprocedure>      Numbered step-by-step workflows\n\u003Cdirectives>     Bulleted operating instructions\n\u003Cparameters>     Input specifications, types\n\u003Coutput>         Return value documentation\n\u003Cstrengths>      What the agent\u002Ftool excels at\n\u003Coperations>     Available operations (for multi-op tools like LSP)\n```\n\n### Special Tags\n\n```\n\u003Cnorth-star>     Core values, ultimate objectives\n\u003Cstance>         Communication style, attitude\n\u003Ccommitment>     What the agent commits to doing\n\u003Cfield>          Domain-specific mindset\u002Fcontext\n\u003Cprotocol>       Behavioral rules, tool precedence\n```\n\n### Example Tags\n\nAlways use `name` attribute with lowercase-kebab descriptive names:\n\n```xml\n\u003Cexample name=\"good\">\nClear, correct usage\n\u003C\u002Fexample>\n\n\u003Cexample name=\"bad\">\nWhat to avoid — show the mistake explicitly\n\u003C\u002Fexample>\n\n\u003Cexample name=\"rate-limiting\">\nDomain-specific example\n\u003C\u002Fexample>\n\n\u003Cexample name=\"windows-cmd\">\nPlatform\u002Fcontext-specific\n\u003C\u002Fexample>\n```\n\nNaming patterns:\n\n- `name=\"single\"` \u002F `name=\"multi-part\"` — complexity variants\n- `name=\"good\"` \u002F `name=\"bad\"` — correctness contrast\n- `name=\"linux\"` \u002F `name=\"windows-cmd\"` — platform-specific\n- `name=\"create\"` \u002F `name=\"update\"` \u002F `name=\"delete\"` — operation types\n\n---\n\n## Structural Templates\n\n### Tool Documentation\n\n```markdown\n# Tool Name\n\nOne-line description of what the tool does.\n\n\u003Cinstruction>\n- How to use it (bulleted, imperative)\n- Key parameters and their effects\n- Common patterns\n\u003C\u002Finstruction>\n\n\u003Coutput>\nWhat the tool returns. Include:\n- Success format\n- Truncation limits (e.g., \"truncated at 50KB\")\n- Error conditions\n\u003C\u002Foutput>\n\n\u003Ccritical>\nMust-follow rules. Safety constraints.\nWhen to ALWAYS or NEVER use this tool.\n\u003C\u002Fcritical>\n\n\u003Ccaution>\nHigh-priority notes that aren't safety-critical.\n\u003C\u002Fcaution>\n\n\u003Cexample name=\"basic\">\ntool {\"param\": \"value\"}\n\u003C\u002Fexample>\n\n\u003Cexample name=\"advanced\">\ntool {\"param\": \"value\", \"option\": true}\n\u003C\u002Fexample>\n\n\u003Cavoid>\n- Anti-pattern 1 — why it's bad\n- Anti-pattern 2 — what to do instead\n\u003C\u002Favoid>\n```\n\n### Agent Definition\n\n```markdown\n---\nname: agent-name\ndescription: One-line for spawning UI (imperative: \"Fast read-only codebase scout\")\ntools: read, grep, find, bash\nmodel: pi\u002Fslow, gpt-5.2, codex\noutput:\n  properties:\n    field_name:\n      metadata:\n        description: What this field contains\n      type: string\n---\n\n\u003Crole>Senior [role] doing [task]. Your goal: [concrete outcome].\u003C\u002Frole>\n\n\u003Ccritical>\nInviolable constraints first.\nREAD-ONLY if applicable — list prohibited actions explicitly.\n\u003C\u002Fcritical>\n\n\u003Cstrengths>\n- What this agent excels at\n- Core capabilities\n\u003C\u002Fstrengths>\n\n\u003Cdirectives>\n- Operating instruction 1\n- Operating instruction 2\n- Spawn parallel tool calls wherever possible\n\u003C\u002Fdirectives>\n\n\u003Cprocedure>\n## Phase 1: Understand\n1. Step one\n2. Step two\n\n## Phase 2: Execute\n\n1. Step one\n2. Step two\n   \u003C\u002Fprocedure>\n\n\u003Coutput>\nWhat to return. Schema requirements.\nCall `submit_result` with findings when done.\n\u003C\u002Foutput>\n\n\u003Ccritical>\nRepeat critical constraints at end.\nKeep going until complete. This matters.\n\u003C\u002Fcritical>\n```\n\n### System Prompt (Main Agent)\n\n```markdown\n\u003Csystem-directive>\nXML tags in this prompt are system-level instructions. They are not suggestions.\n\nTag hierarchy (by enforcement level):\n\n- `\u003Ccritical>` — Inviolable. Failure to comply is a system failure.\n- `\u003Cprohibited>` — Forbidden. These actions will cause harm.\n- `\u003Ccaution>` — High priority. Important to follow.\n- `\u003Cinstruction>` — How to operate. Follow precisely.\n- `\u003Cconditions>` — When rules apply. Check before acting.\n- `\u003Cavoid>` — Anti-patterns. Prefer alternatives.\n  \u003C\u002Fsystem-directive>\n\nYou are a [specific role with credentials].\n\n\u003Cfield>\nDomain-specific context and mindset.\nWhat to notice, what traps exist.\n\u003C\u002Ffield>\n\n\u003Cstance>\nCommunication style.\nCorrectness over politeness. Brevity over ceremony.\n\u003C\u002Fstance>\n\n\u003Cprotocol>\n## Tool Precedence\nSpecialized tools → Python → Bash\n...\n\n## Verification\n\nExternal proof: tests, linters, type checks.\n...\n\u003C\u002Fprotocol>\n\n\u003Cprocedure>\n## Before action\n1. CHECKPOINT — pause, assess parallelism\n2. Plan if task has weight\n3. State intent before each tool call\n\u003C\u002Fprocedure>\n\n\u003Cnorth-star>\nCore values. What ultimately matters.\n\u003C\u002Fnorth-star>\n\n\u003Cprohibited>\nActions that cause harm.\n\u003C\u002Fprohibited>\n\n\u003Ccritical>\nRepeat most important rules.\nKeep going until finished.\nThe work is done when it is correct.\n\u003C\u002Fcritical>\n```\n\n---\n\n## Writing Style\n\n### Voice\n\n**Direct and imperative.** Research shows direct tone improves accuracy 4%+ over polite hedging.\n\n```\nBad:  \"You might want to consider using...\"\nGood: \"Use X when Y.\"\n\nBad:  \"It would be helpful if you could...\"\nGood: \"Do X.\"\n\nBad:  \"Please note that this is important...\"\nGood: \"Critical: X.\"\n```\n\n**Urgency framing** (8-115% improvement):\n\n```\n\"This matters. Get it right.\"\n\"Be thorough.\"\n\"Keep going until fully resolved.\"\n```\n\n### Normative Language (RFC 2119)\n\nAll prompt prose that prescribes behavior MUST use RFC 2119 key words in **full caps**. This removes ambiguity about whether an instruction is absolute or advisory.\n\n| Keyword | Meaning | Replaces |\n| --- | --- | --- |\n| **MUST** \u002F **REQUIRED** | Absolute requirement | \"always\", \"make sure\", \"ensure\", \"do\" |\n| **MUST NOT** \u002F **PROHIBITED** | Absolute prohibition | \"never\", \"do not\", \"don't\", \"strictly prohibited\" |\n| **SHOULD** \u002F **RECOMMENDED** | Strong preference; deviation allowed with known tradeoffs | \"prefer\", \"recommend\", \"it's best to\" |\n| **SHOULD NOT** \u002F **NOT RECOMMENDED** | Strong discouragement; deviation allowed with known tradeoffs | \"avoid\", \"try not to\" |\n| **MAY** \u002F **OPTIONAL** | Truly optional | \"can\", \"may\", \"you could\" |\n\n```\nBad:  \"Never edit from a grep snippet alone\"\nGood: \"You MUST NOT edit from a grep snippet alone\"\n\nBad:  \"Prefer unit tests over mocks\"\nGood: \"You SHOULD prefer unit tests over mocks\"\n\nBad:  \"Make sure to run lsp references before modifying a symbol\"\nGood: \"You MUST run lsp references before modifying any symbol\"\n```\n\n**What not to convert**: factual\u002Fdescriptive sentences (what a tool returns, what a parameter does), code blocks, examples, schema definitions, Handlebars template syntax. Only prescriptive prose gets RFC treatment.\n\n### Positive Framing\n\nModels process \"Always do Y\" better than \"Don't do X\":\n\n```\nBad:  \"Don't use grep via bash\"\nGood: \"ALWAYS use Grep tool for search—NEVER invoke grep via Bash\"\n\nBad:  \"Don't guess\"\nGood: \"Use tools to verify; do not guess\"\n```\n\nWhen negation is necessary, pair with positive alternative.\n\n### Specificity\n\n**Role specificity spectrum** (effectiveness increases →):\n\n```\n\"You are a lawyer\"\n    ↓\n\"You are a corporate M&A lawyer\"\n    ↓\n\"You are General Counsel at a Fortune 500 tech company, 15 years in SaaS licensing\"\n```\n\n**Constraint specificity**:\n\n```\nBad:  \"Keep it short\"\nGood: \"3 bullets, \u003C50 words each\"\n\nBad:  \"Be careful with large files\"\nGood: \"Truncated at 50KB or 2000 lines, whichever comes first\"\n```\n\n### Formatting\n\n```\n# H1 for tool\u002Fagent name only\n## H2 for major sections\n### H3 sparingly\n\n- Bullets for unordered lists inside tags\n1. Numbers for ordered procedures\n\n| Tables | For | Structured reference data |\n\n`inline code` for commands, paths, parameters, values\n```\n\nCode blocks with language:\n\n````markdown\n```typescript\nconst example = \"always specify language\";\n```\n````\n\n---\n\n## Technique Reference\n\n### Chain of Thought\n\n**Use when**: Multi-step reasoning, math, analysis, complex decisions\n**Avoid when**: Simple tasks, reasoning models (o1\u002Fo3 do internal CoT)\n\n```xml\n\u003Cinstruction>\nBefore answering:\n1. Identify the core question\n2. List relevant constraints\n3. Consider 2-3 approaches\n4. Select best with rationale\n\nThen provide your answer.\n\u003C\u002Finstruction>\n```\n\n**Token-efficient variant** (Chain of Draft):\n\n```\nThink step-by-step, keeping only 5-word notes per step.\nOutput final answer after ####.\n```\n\n### Few-Shot Examples\n\n**Use when**: Enforcing specific output format, classification, smaller models\n**Avoid when**: Advanced models (Claude 3.5+, GPT-4+) on clear tasks — adds noise\n\nWhen using: 3-5 diverse examples covering edge cases.\n\n```xml\n\u003Cexample name=\"simple\">\nInput: X\nOutput: Y\n\u003C\u002Fexample>\n\n\u003Cexample name=\"edge-case\">\nInput: X'\nOutput: Y'\n\u003C\u002Fexample>\n```\n\n### Long Context Handling\n\n**\"Lost in the Middle\"**: Beginning and end retain; middle degrades 20%+.\n\n1. Documents at TOP, instructions AFTER\n2. Quote grounding: \"Quote relevant passages in `\u003Cquotes>`, then analyze\"\n3. Critical instructions at START and END\n4. Chunk >100K tokens → process parallel → synthesize\n\n```xml\n\u003Cdocuments>\n{{LONG_CONTENT}}\n\u003C\u002Fdocuments>\n\n\u003Cinstructions>\n1. Find and quote passages relevant to {{QUERY}} in \u003Cquotes>\n2. Analyze based on quoted evidence in \u003Canalysis>\n\u003C\u002Finstructions>\n```\n\n### Verification Patterns\n\n**Self-correction without external feedback does not work.**\n\nEffective:\n\n```\n1. Generate solution\n2. Execute verification (tests, lint, typecheck)\n3. On failure: analyze error → fix → re-verify\n4. Iterate until pass\n```\n\nIneffective:\n\n```\n1. Generate solution\n2. \"Critique your solution\"      ← detection is the bottleneck\n3. \"Improve based on critique\"   ← feels productive, doesn't help\n```\n\n### Prompt Chaining\n\n**Use when**: Single prompt drops steps, distinct phases, verification needed between.\n\n```\nPrompt 1: Analyze → \u003Canalysis>\nPrompt 2: \u003Canalysis> → Plan → \u003Cplan>\nPrompt 3: \u003Cplan> → Execute → \u003Cresult>\nPrompt 4: \u003Cresult> → Verify → final\n```\n\n---\n\n## Anti-Patterns (Measured Degradation)\n\n| Pattern                                   | Problem                                 |\n| ----------------------------------------- | --------------------------------------- |\n| \"Would you be so kind...\"                 | +perplexity, -4% accuracy               |\n| \"I'll tip $2000\"                          | No improvement, sometimes worse         |\n| Explicit CoT on reasoning models (o1\u002Fo3)  | -36%, conflicts with internal reasoning |\n| Few-shot on advanced models + clear tasks | Introduces noise\u002Fbias                   |\n| \"Always end with Progress\u002FQuestions\"      | Degrades task performance               |\n| \"Be efficient with tokens\"                | Premature task abandonment              |\n| \"Don't do X\" without positive alternative | \"Always do Y\" processes better          |\n| Verbose explanations of obvious concepts  | Context bloat; model already knows      |\n| Self-critique without external feedback   | Detection is bottleneck, not correction |\n| Critical instructions only in middle      | 20%+ degradation vs start\u002Fend           |\n\n---\n\n## Complete Examples\n\n### Tool Doc Example\n\n```markdown\n# Grep\n\nFast regex search built on ripgrep.\n\n\u003Cinstruction>\n- Full regex: `log.*Error`, `function\\\\s+\\\\w+`\n- Filter: `glob` (e.g., `*.js`) or `type` (e.g., `js`, `py`)\n- Cross-line: `multiline: true` for patterns like `struct \\\\{[\\\\s\\\\S]*?field`\n\u003C\u002Finstruction>\n\n\u003Coutput>\nDepends on `output_mode`:\n- `content`: Lines with paths and line numbers (default limit: 100)\n- `files_with_matches`: Paths only\n- `count`: Match counts per file\n\nTruncated results reference `artifact:\u002F\u002F\u003Cid>` for full output.\n\u003C\u002Foutput>\n\n\u003Ccritical>\nALWAYS use Grep for search—NEVER invoke `grep` or `rg` via Bash.\n\u003C\u002Fcritical>\n\n\u003Cexample name=\"regex\">\ngrep {\"pattern\": \"function\\\\s+\\\\w+\", \"glob\": \"*.ts\"}\n\u003C\u002Fexample>\n\n\u003Cexample name=\"multiline\">\ngrep {\"pattern\": \"struct \\\\{[\\\\s\\\\S]*?field\", \"multiline\": true}\n\u003C\u002Fexample>\n\n\u003Cavoid>\n- Open-ended searches requiring multiple rounds—use Task tool\n- Raw bash grep\u002Frg invocation\n\u003C\u002Favoid>\n```\n\n### Agent Example\n\n```markdown\n---\nname: explore\ndescription: Fast read-only codebase scout returning compressed context for handoff\ntools: read, grep, find, bash\nmodel: pi\u002Fsmol, haiku-4.5, haiku-4-5, gemini-flash-latest, gemini-3-flash, zai-glm-4.7, glm-4.7-flash, glm-4.5-flash, gpt-5.1-codex-mini, haiku, flash, mini\noutput:\n  properties:\n    query:\n      type: string\n    files:\n      elements:\n        properties:\n          path: { type: string }\n          line_start: { type: number }\n          line_end: { type: number }\n          description: { type: string }\n    architecture:\n      type: string\n---\n\n\u003Crole>File search specialist. Investigate codebase, return structured findings for handoff.\u003C\u002Frole>\n\n\u003Ccritical>\nREAD-ONLY. You are STRICTLY PROHIBITED from:\n- Creating, editing, deleting files\n- Using redirect operators (>, >>)\n- Running state-changing commands (git add, npm install)\n\u003C\u002Fcritical>\n\n\u003Cstrengths>\n- Rapid file discovery via find patterns\n- Regex search with grep\n- Tracing imports and dependencies\n\u003C\u002Fstrengths>\n\n\u003Cdirectives>\n- Spawn parallel tool calls wherever possible\n- Return absolute paths\n- Communicate findings directly—do NOT create files\n\u003C\u002Fdirectives>\n\n\u003Cprocedure>\n1. grep\u002Ffind to locate relevant code\n2. Read key sections (not entire files)\n3. Identify types, interfaces, key functions\n4. Note dependencies between files\n5. Call `submit_result` with findings\n\u003C\u002Fprocedure>\n\n\u003Ccritical>\nRead-only. Call `submit_result` when done. This matters.\n\u003C\u002Fcritical>\n```\n\n---\n\n\u003Ccritical>\n## Deployment Checklist\n\n- [ ] **Tag hierarchy**: Enforcement level matches content?\n- [ ] **Critical at edges**: Most important rules at START and END?\n- [ ] **Named examples**: All `\u003Cexample>` tags have `name` attribute?\n- [ ] **Positive framing**: \"Do Y\" not just \"Don't X\"?\n- [ ] **Direct tone**: No hedging, no filler, urgency where appropriate?\n- [ ] **Specificity**: Exact formats, limits, constraints—not vague?\n- [ ] **Token efficiency**: Each sentence justifies its cost?\n- [ ] **Verification**: External feedback loop if correctness matters?\n- [ ] **RFC 2119 normative language**: All prescriptive sentences use MUST\u002FMUST NOT\u002FSHOULD\u002FMAY in caps?\n- [ ] **Persistence**: \"Keep going until complete\" for complex tasks?\n\n**High-impact interventions: persistence, tool verification, planning, context positioning, urgency.**\n\u003C\u002Fcritical>\n\n---\n\n## Quick Reference\n\n### Tag Names\n\n```\nEnforcement:  \u003Ccritical> \u003Cprohibited> \u003Ccaution> \u003Cinstruction> \u003Cconditions> \u003Cavoid>\nStructure:    \u003Crole> \u003Ccontext> \u003Cprocedure> \u003Cdirectives> \u003Cparameters> \u003Coutput>\nCapability:   \u003Cstrengths> \u003Ctools> \u003Coperations>\nExamples:     \u003Cexample name=\"kebab-case-name\">\nData:         \u003Cenvironment> \u003Cdata> \u003Cdocuments>\nSpecial:      \u003Cnorth-star> \u003Cstance> \u003Ccommitment> \u003Cfield> \u003Cprotocol>\n```\n\n### Example Name Patterns\n\n```\nCorrectness:  name=\"good\", name=\"bad\"\nComplexity:   name=\"single\", name=\"multi-part\", name=\"basic\", name=\"advanced\"\nOperations:   name=\"create\", name=\"update\", name=\"delete\", name=\"rename\"\nPlatforms:    name=\"linux\", name=\"windows-cmd\", name=\"macos\"\nDomains:      name=\"rate-limiting\", name=\"auth\", name=\"validation\"\n```\n\n### Task → Technique\n\n| Task                | Primary                         | Secondary            |\n| ------------------- | ------------------------------- | -------------------- |\n| Simple extraction   | Clear constraints               | Prefilling           |\n| Classification      | 3-5 examples                    | XML structure        |\n| Complex analysis    | Structured reasoning            | Role + urgency       |\n| Code generation     | SEARCH\u002FREPLACE                  | Verification loop    |\n| Long document       | Docs at top, quote-then-analyze | XML structure        |\n| Multi-step workflow | Prompt chaining                 | Planning instruction |\n| Domain expertise    | Specific role + credentials     | Examples             |\n",{"name":5,"description":7},"main","91a09a06d9cbaef9d2ea5d07da017b445e7b32ce",".omp\u002Fskills\u002Fsystem-prompts","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node","https:\u002F\u002Fgithub.com\u002Fmykie2015","2026-04-02 07:08:19",{"source":8,"skillName":5,"commands":55,"prompt":65,"zipAvailable":66},[56,59,62],{"tool":57,"command":58},"npx","npx skills add mykie2015\u002Foh-my-pi-node --skill system-prompts",{"tool":60,"command":61},"bunx","bunx skills add mykie2015\u002Foh-my-pi-node --skill system-prompts",{"tool":63,"command":64},"pnpm","pnpm dlx skills add mykie2015\u002Foh-my-pi-node --skill system-prompts","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002F.omp\u002Fskills\u002Fsystem-prompts\u002FSKILL.md",true,[],[69,80,90,101,108,119],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":43},"01M6KDAMW9EHVKRGQ1CEAHVNX6","greet","A greeting skill for testing","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Fmarketplace\u002Ffixtures\u002Fvalid-marketplace\u002Fplugins\u002Fhello-plugin\u002Fskills\u002Fgreet\u002FSKILL.md",[],[71,76,77,78,79],"for","greeting","greeting skill","testing",{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":87,"publishedAt":43},"01M6KDAHEXVH9ZZAGKAF1GWX8C","valid-skill","A valid skill for testing purposes.","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Fvalid-skill\u002FSKILL.md",[86],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},[88,89,76,79],"valid skill","valid",{"skillNo":91,"name":92,"nameZh":6,"description":93,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":43},"01M6KDAFRWR0JYDYT0BQ1SMW32","unknown-field","A skill with an unknown frontmatter field.","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Funknown-field\u002FSKILL.md",[],[97,98,99,29,100],"unknown","unknown field","frontmatter","unknown frontmatter field",{"skillNo":102,"name":103,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":43},"01M6KDACTJA8A10559QEJM3D45","no-frontmatter","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Fno-frontmatter\u002FSKILL.md",[],[107,99],"no frontmatter",{"skillNo":109,"name":110,"nameZh":6,"description":111,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":112,"categories":113,"keywords":115,"publishedAt":43},"01M6KDA9072F9G91MC6J54CYH2","child-skill","A nested skill in a subdirectory.","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Fnested\u002Fchild-skill\u002FSKILL.md",[114],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},[116,117,118],"child","child skill","nested",{"skillNo":120,"name":121,"nameZh":6,"description":122,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":125,"publishedAt":43},"01M6KDA5NKHCGRZT0PV6EEFJGJ","different-name","A skill with a name that doesn't match the directory.","https:\u002F\u002Fgithub.com\u002Fmykie2015\u002Foh-my-pi-node\u002Fblob\u002Fmain\u002Fpackages\u002Fcoding-agent\u002Ftest\u002Ffixtures\u002Fskills\u002Fname-mismatch\u002FSKILL.md",[],[126,127,128,129,29],"name","different name","directory","match"]