| name | skill-builder-guide |
| description | AI Agent Skills creation methodology and execution pipeline. Triggered when creating skill systems for projects, generating project-specific skills, understanding skill templates, model delegation, or validating skill accuracy. This is a META skill — it creates other skills as its output. |
| model_tier | L1 |
| skill_tier | meta |
| composes | [{"planning":"delegation"},{"functional":"change-model"},{"atomic":"example-dev"},{"atomic":"example-code-map"},{"atomic":"example-delegation"}] |
| composed_by | [] |
| context_budget | {"l1_metadata":120,"l2_body":4000,"l3_references":20000} |
| version | 2.0.0 |
| status | active |
| review_by | "2026-07-28T00:00:00.000Z" |
| trust_level | internal |
| requires_network | false |
| requires_file_write | true |
| compatibility | requires git, python3, pyyaml |
| allowed_tools | Bash Read Write Edit Grep Glob Agent |
| evolution | {"usage_count":0,"last_corrections":[],"stale_markers":[]} |
Skill Builder Guide — Meta-Skill for Skill Creation
Position: meta tier skill — output is not business code, but new skill files (SKILL.md + openai.yaml).
Core capability: Analyze target project → scan code to extract patterns → generate project-specific skills from templates → validate accuracy → user confirmation.
Composition: Orchestrates code-map, dev, delegation as atomic supports.
Full spec: SKILL-BUILDER-GUIDE.md (project root) · Quick start: docs/quick-start.md
Trigger Conditions
- Creating a skill system for a project
- "create skill" / "generate project-specific skill" / "skill template"
- "model tier" / "L0 delegation" / "skill validation"
- "skill packaging" / "skill directory" / "skill builder"
- "openai.yaml" / "frontmatter spec"
Related Skills
1. Skill System Architecture
Every skill is defined on two independent dimensions, orthogonal to each other:
| Execution Axis (model_tier) | Composition Axis (skill_tier) |
|---|
| Question | Who executes? | Where in the composition graph? |
| Values | L0/L1/L2/L3 | meta/planning/functional/atomic |
Execution axis: L0=Haiku mechanical · L1=Sonnet bounded implementation · L2=Sonnet/Opus multi-step reasoning · L3=Opus architectural decisions
Composition axis: meta=creates skills · planning=task decomposition & routing · functional=reusable multi-step routines · atomic=single information source
atomic+L0 = pure lookup (Haiku); functional+L1 = multi-step reasoning (Sonnet)
2. Skill Type Decision
| Project Profile | Count | Skill Combo |
|---|
| Small | 2 | standards + workflow |
| Medium | 3 | + code-map |
| Large | 4 | + change-model |
| Complex | 5-6 | + call-chain + scripts + delegation |
Default model tier quick reference:
| Skill Type | Execution | Composition |
|---|
| Standards (dev) | L1 | atomic |
| Code Map (code-map) | L0 | atomic |
| Workflow (workflow) | L1 | functional |
| Scripts (scripts) | L0 | atomic |
| Call-Chain (call-chain) | L1 | functional |
| Change Model (change-model) | L1 | functional |
| Delegation (delegation) | L1 | planning |
Full selection matrix: references/skill-types-catalog.md.
3. Five-Phase Execution Pipeline
Phase 1: ANALYZE ──→ Phase 2: SCAN ──→ Phase 3: GENERATE ──→ Phase 4: VALIDATE ──→ Phase 5: CONFIRM
(L1-Sonnet) (L0-Haiku) (L1-Sonnet) (L0-Haiku) (L1-Sonnet)
│ │ │ │ │
Skill plan Code samples Skill files Validation report approved/revise
Phase 1: ANALYZE — Project Analysis
- Executor: Sonnet (L1)
- Output:
{ skill_plan: [{ name, skill_tier, model_tier, reason }] }
Analyze project type, tech stack, module structure, team size. Output plan in Conclusion/Basis/Uncertainty format.
Gate: Phase 1 must pause after completion. Wait for user confirmation of the skill plan. Do NOT proceed to Phase 2 without approval.
Phase 2: SCAN — Code Scanning
- Executor: Haiku (L0) — must delegate
- Output: Code samples per layer
| Layer | Scan Target | Record |
|---|
| API Layer | Route declarations, param validation, response wrapping | {method}, {class/function name} |
| Service Layer | Abstract interfaces, consistency mgmt, param conversion | {yes/no}, {method} |
| Data Layer | ORM approach, query organization, pagination | {framework}, {method} |
| Integration Layer | Remote calls, message queues, scheduled tasks | {list}, {config} |
Detailed scan dimensions: references/pipeline/phase-2-scan.md.
Phase 3: GENERATE — Generate Skill Files
- Executor: Sonnet (L1)
- Output:
SKILL.md + agents/openai.yaml → default .claude/skills/{name}/ (template projects use skills/{name}/)
Hard rules:
- Fill frontmatter per references/frontmatter-spec.md, include dual-axis fields
- Extract trigger words from project code (5-15, action + query)
- Fill actual version numbers from scan — no placeholders
- Use actual scanned code snippets (sanitized)
- SKILL.md body ≤5000 tokens
Generated skills default to concise, accurate English content.
Detailed prompt template: references/pipeline/phase-3-generate.md.
Phase 4: VALIDATE — Verify
- Executor: Haiku (L0) — must delegate
python scripts/validate-skills.py .claude/skills/{skill-name}
python scripts/validate-skills.py skills/{skill-name}
python scripts/validate-skills.py .claude/skills/{skill-name} --semantic
| Layer | Pass Condition |
|---|
| V1 Format | frontmatter complete, triggers ≥5, YAML valid |
| V2 Structure | Dual-axis consistent, composition graph closed |
| V3 Semantic | File paths ≥95%, method names ≥90%, version numbers 100% |
Below standard = must not publish. See references/validation-protocol.md.
Phase 5: CONFIRM — User Confirmation
- Executor: Sonnet (L1)
- Output:
approved | revise({feedback}) | reject
Confirm: version numbers correct, standard patterns match expectations, compatibility patterns accurately explained, user docs merged (user docs take precedence).
4. Directory Structure
| Path | Use Case | Auto-Load | Slash Cmd |
|---|
.claude/skills/{name}/ | Production project skills | ✅ | ✅ |
skills/{name}/ | Template/methodology reference | ❌ | ❌ |
~/.claude/skills/{name}/ | Global cross-project skills | ✅ | ✅ |
{skill-name}/
├── SKILL.md # L2: Core instruction body (≤5000 tokens)
├── agents/
│ └── openai.yaml # L1: Trigger config
├── references/ # L3: Deep reference, loaded on demand
├── scripts/ # L4: Executable scripts
└── assets/ # Static resources
CLAUDE.md integration: Routing table optional with .claude/skills/, required with skills/. Mandatory delegation rules always at top.
5. Core Spec Reference
6. Model Tier
L1 — Sonnet / meta tier: Requires understanding project structure, extracting code patterns, generating structured skill docs from templates, iterating based on validation feedback. Scan sub-tasks delegate to L0 — Haiku.