skill-builder（技能构建）

Lev Skill Builder

Overview

Unified skill creation hub. Routes between documentation codification (Skill_Seekers v2.7.4) and new skill authoring (skill-creator standards). Supports website scraping, GitHub repository analysis (with AST parsing), PDF extraction, and unified multi-source with conflict detection.

When to Use This Skill

Use skill-builder when you need to:

Do NOT use for: editing existing skills (use editor), searching skills (use lev get), or installing skills (use clawdhub).

Quick Decision Tree

What does the user want?
│
├─→ Convert existing docs to skill?
│   ├─→ Website only? → Website Scraping
│   ├─→ GitHub repo only? → GitHub Analysis
│   ├─→ PDF only? → PDF Extraction
│   └─→ Multiple sources? → Unified Multi-Source
│
├─→ Create NEW skill from scratch?
│   └─→ Route to skill-creator
│       - SKILL.md format: YAML frontmatter + markdown body
│       - Progressive disclosure: metadata → body → references
│       - Body <500 lines, description is PRIMARY trigger
│       - See skill-creator skill for full standards
│
├─→ Combine/merge existing skills?
│   └─→ Route to BOTH:
│       1. Analyze existing skills (codifier patterns)
│       2. Author merged skill (skill-creator standards)
│       3. Ensure router pattern if subsumes others
│
├─→ Export for non-Claude platform?
│   └─→ Package with --target (gemini|openai|markdown)
│
└─→ First time? → references/setup.md

Skill Installation from External Sources

When intake routes a skills.sh URL or skill:// to skill-builder:

Step 1: Acquire → skills-db staging

# Extract GitHub repo URL from skills.sh page (WebFetch for the link ONLY)
git clone --depth 1 {repo} /tmp/skill-intake-{ts}/
# Find the skill: find /tmp/skill-intake-{ts}/ -name "SKILL.md" -path "*{name}*"
# Stage to skills-db (NOT directly to active):
cp -r /tmp/skill-intake-{ts}/skills/{name}/ ~/.agents/skills-db/_workshop/{source}/{name}/
rm -rf /tmp/skill-intake-{ts}/

Step 2: Validate (HARD GATES — run on staged copy)

file=~/.agents/skills-db/_workshop/{source}/{name}/SKILL.md
head -1 "$file" | grep -q "^---$"           # Has YAML frontmatter
grep -q "^name:" "$file"                     # Has name field
grep -q "^description:" "$file"              # Has description field
awk '/^---$/{c++} c==2{exit}' "$file"        # Has closing ---

If any hard gate fails → REJECT. Move to .archive/ or delete. Do NOT promote.

Step 2b: Security Audit (external skills ONLY — skip for local authoring)

Three sequential scanners with early termination. Each scanner is progressively more expensive. See references/security-audit-gates.md for full scoring rubric and quarantine protocol.

Scanner 1: Structural Decompile (free, instant)

python3 ~/.agents/skills-db/security/skill-decompile/decompile.py "$file" --output yaml
# Check: risk_score < 60 to proceed
# If risk_score >= 60: REJECT with flags

Scanner 2: Semantic Scan (agent-based, ~10s) Load security-scanner skill (hiroro) against the staged skill directory:

# Runs in sandboxed subagent with Read-only tools
# Checks for malicious NL instructions, social engineering patterns
# Output: pass/warn/fail

Scanner 3: AgentShield CLI (CLI-based, ~5s)

npx ecc-agentshield scan --path "$staged_dir" --min-severity medium
# Checks: hook injection, MCP risks, overpermissive configs
# Output: findings JSON

Security Verdict:

Step 3: Prior Art Check

• Search ~/.agents/skills/ for name/trigger overlap
• Exact match → compare quality (line count, scoring), keep better version
• Partial overlap (>30% triggers) → recommend merge
• No overlap → proceed to scoring

Step 4: Quality Score (run on staged copy, BEFORE promotion)

Score 1-10 on 5 dimensions (read the SKILL.md content):

Grades: A (8+) promote, B (7-7.9) promote with note, C (5-6.9) hold in _todo, D (<5) reject

Step 5: Catalog (move from staging to skills-db home)

# A/B grade → catalog in skills-db (DEFAULT destination):
mv ~/.agents/skills-db/_workshop/{source}/{name}/ ~/.agents/skills-db/{domain}/{name}/
# C grade → move to _todo for enhancement:
mv ~/.agents/skills-db/_workshop/{source}/{name}/ ~/.agents/skills-db/_todo/{name}/
# D grade → reject:
mv ~/.agents/skills-db/_workshop/{source}/{name}/ ~/.agents/skills-db/.archive/{name}/

Step 5b: Activate (ONLY if user requests)

# Only when user explicitly wants the skill loaded into Claude Code:
cp -r ~/.agents/skills-db/{domain}/{name}/ ~/.agents/skills/{name}/
# Or symlink:
ln -s ~/.agents/skills-db/{domain}/{name}/ ~/.agents/skills/{name}

NOTE: Most skills stay in skills-db. Only day-to-day/global operational skills get activated. The user decides when to activate — skill-builder should PROPOSE activation, not assume it.

Step 6: Lifecycle Check

• Is this skill a candidate for merging into an existing hub/router?
• Does it overlap with 2+ existing skills in the same domain?
• If yes → recommend leaf→hub→router graduation

Skill Lifecycle: Leaf → Hub → Router

Skills grow organically through 3 stages:

Growth triggers

• Merge: 2+ skills share >30% intent overlap → merge into hub
• Router: Merged hub exceeds ~400L → graduate to router (move content to sub-skills)
• Split: Single skill >500L without references/ → split into skill + references/

Router graduation checklist

1. Create {domain}-hub/ directory
2. Write SKILL.md routing header (<100L): decision tree + sub-skill table
3. Add skill_type: router and subsumes: [list] to frontmatter
4. Keep sub-skills as independent SKILL.md files
5. Router description MUST include ALL sub-skill triggers (union of tags)

Full Pipeline (The Correct Order)

Every codification job follows this pipeline. Do NOT skip steps.

1. PRIOR ART CHECK → Does this skill already exist?
   ├─ lev get "{name}" --scope=knowledge --pattern="SKILL.md"
   ├─ grep -rl "{name}" ~/.claude/skills/
   └─ If found:
      • Exact match → use as-is or enhance existing, skip to step 4
      • Partial overlap → recommend merge/consolidation with existing
      • Related but different → proceed, note in description for routing

2. ESTIMATE (websites only) → How big is this job?
   └─ skill-seekers estimate configs/{name}.json
      • < 5K pages: single skill, proceed normally
      • 5K-10K: consider category split
      • 10K+: must split with router strategy

3. EXTRACT → Get the raw content
   ├─ Website:  skill-seekers scrape --name {name} --url {url}
   ├─ GitHub:   skill-seekers github --repo {owner/repo}
   ├─ PDF:      skill-seekers pdf --pdf {file} --name {name}
   └─ Unified:  skill-seekers unified --config {config.json}

4. ENHANCE → Transform raw extraction into a real skill
   └─ skill-seekers enhance output/{name}/

   ⚠️ KNOWN BUG (skill-seekers ≤2.7.4): enhance passes file path
   as positional arg to claude CLI. SKILL.md never gets updated.
   USE WORKAROUND: scripts/enhance-workaround.sh output/{name}

5. REVIEW → Check the enhanced output
   • Compare frontmatter against standards (what/how/when/why + triggers)
   • Verify description is the PRIMARY trigger mechanism
   • Check progressive disclosure (SKILL.md <500 lines, rest in references/)
   • Validate: no stats-only wrappers, real actionable content

6. PACKAGE → Bundle for distribution
   └─ echo "y" | skill-seekers package output/{name}/ [--target claude|gemini|openai|markdown]

7. INSTALL → Put it where it belongs
   ├─ Global:  cp -r output/{name} ~/.claude/skills/{name}/
   ├─ Project: cp -r output/{name} .claude/skills/{name}/
   └─ Upload:  output/{name}.zip → https://claude.ai/skills

Quick Reference: Essential Commands

1. Installation Check

command -v skill-seekers || python3 -m skill_seekers --version

If not installed, see references/setup.md for uv/venv/pip auto-detection.

2. Website Scraping (Preset)

skill-seekers scrape --config configs/react.json --enhance-local
skill-seekers package output/react/

3. Website Scraping (Custom)

skill-seekers scrape --name myframework --url https://docs.example.com/
skill-seekers package output/myframework/

4. GitHub Repository Analysis

skill-seekers github --repo facebook/react
skill-seekers package output/react/

Supports deep AST parsing for Python, JavaScript, TypeScript, Java, C++, Go. Extracts APIs, issues, PRs, changelogs, and detects doc-vs-code conflicts.

5. PDF Extraction

skill-seekers pdf --pdf docs/manual.pdf --name myskill
skill-seekers package output/myskill/

Supports OCR for scanned PDFs, table extraction, password-protected files, and parallel processing.

6. Unified Multi-Source (Docs + GitHub + PDF)

skill-seekers unified --config configs/react_unified.json
skill-seekers package output/react/

Combines all sources with automatic conflict detection and intelligent merging.

7. Multi-Platform Export

# Claude (default)
skill-seekers package output/react/

# Google Gemini
skill-seekers package output/react/ --target gemini

# OpenAI ChatGPT
skill-seekers package output/react/ --target openai

# Generic Markdown (any LLM)
skill-seekers package output/react/ --target markdown

For large docs (10K+ pages), async mode, three-stream GitHub analysis, splitting strategies, and troubleshooting, see references/advanced-commands.md.

Frontmatter Design Philosophy

The Description is Everything

The description field is the PRIMARY trigger mechanism. Agents read descriptions to route. Body loads AFTER triggering.

Every description needs What/How/When/Why + tag soup:

description: |
  [WHAT] - What this skill does (1 sentence)
  [HOW] - How it works (mechanism, tools used)
  [WHEN] - When to use it (trigger conditions)
  [WHY] - Why it exists (problem it solves)

  Triggers: "tag1", "tag2", "tag3", ...

Intentional Tag Overlap

Skills SHOULD have overlapping triggers for semantic routing:

Anti-pattern: Descriptions that only say WHAT without WHEN triggers:

# BAD
description: Manages memory storage and retrieval.

# GOOD
description: |
  Store and recall information across sessions using AutoMem.
  Triggers: "remember", "recall", "forget", "save this", "memory"

The 3-Tool Agent Model

Skills wrap 3 fundamental operations:

1. GET (context gathering) - lev get, lev-research, read
2. EXECUTE (run things) - ralph, bash, Task tool, skills
3. PATCH (modify the graph) - bd, edit, write, memory_store

Categorize your skill's operations to help semantic routing.

Skill Creator Standards (Quick Reference)

When authoring a new skill from scratch:

Structure:

skill-name/
├── SKILL.md (required, 300-500 lines)
│   ├── YAML frontmatter (name, description required)
│   └── Markdown body with inline templates
├── references/ (optional, for API docs or large schemas only)
└── scripts/ (optional, deterministic code)

Rules:

• description is the PRIMARY trigger - include what + when
• Use imperative/infinitive form
• No separate README.md or CHANGELOG.md - the skill IS the docs

Skill Authoring Principles (Hard-Won)

These principles come from rebuilding skills that agents couldn't follow. Apply them to every skill you create.

1. Steps are verbs, not states. Section titles tell the agent what to DO. "Save the handoff" not "EMIT". "Search these 10 sources" not "DISCOVER". If the agent reads the title and doesn't know its next action, the title is wrong.

2. Inline templates at each step. If a step produces an artifact, show the skeleton right there — 10-50 lines. Don't write "load templates/handoff.md" — agents won't do it. The template must be visible where the agent needs it.

3. The first step is the most important output. Whatever the skill's primary deliverable is, that's Step 1. Not background theory. Not architecture diagrams. Not "On Load" setup. The thing the agent produces first.

4. References don't get loaded. Agents don't cat references/gates.md in practice. If information matters for execution, it lives in SKILL.md. Use references/ only for large API docs or schemas that are genuinely loaded on demand via explicit instructions. Never put operational instructions in references/.

5. Prose is the enemy. Every paragraph explaining WHY a rule exists is a paragraph the agent skips to find WHAT to do. Lead with the format/action. Add rationale as a single-line comment if needed. Cut everything else.

6. Labels don't enforce. Marking 14 sections "Hard Contract" is the same as marking zero. The contract IS the inline format. Show the format. Skip the label.

7. Target: 300-500 lines. Under 300 and you're probably missing inline templates. Over 500 and the agent can't hold it in working memory. The 100-150 line "thin SKILL.md + references/" model produces skills that agents can't execute because the operational detail is in files they never read.

Validation (Post-Authoring)

VALIDATION RULES:
1. Can an agent read Step 1's title and immediately act? If not, rewrite.
2. Does every step that produces an artifact have an inline template? If not, add it.
3. Is SKILL.md 300-500 lines? Under → missing templates. Over → cut prose.
4. Are there references/ the agent MUST read to function? Move that content into SKILL.md.
5. Zero sections titled "Hard Contract" or abstract state names (DISCOVER, EMIT, etc.)
6. Total skill > 2000 lines? → Consider sub-skills with router

IDEAL STRUCTURE:
skill-name/
├── SKILL.md (300-500 lines)
│   ├── Frontmatter (name + description with triggers)
│   ├── Steps 1-N (verb-first titles, inline templates)
│   └── Support sections (error handling, routing table)
├── references/ (optional — API docs, large schemas ONLY)
└── scripts/ (optional — deterministic code)

Install Location

Ask user where to install:

• ~/.claude/skills/X/ - User-level (global)
• .claude/skills/X/ - Project-level (local)

Output & Delivery

All workflows create .zip in output/<name>.zip. Users upload to Claude at https://claude.ai/skills. For other platforms, use --target flag.

Working with This Skill

Beginners

1. Run installation check (command -v skill-seekers)
2. If not installed, follow references/setup.md (auto-detects uv/venv/pip)
3. Start with a preset: skill-seekers scrape --config configs/react.json --enhance-local
4. Package: skill-seekers package output/react/

Intermediate

• Use unified multi-source for comprehensive skills (docs + code + PDF)
• Enable async mode for docs over 500 pages
• Use --enhance-local (no API key) or --enhance (with ANTHROPIC_API_KEY) for AI improvement
• Export to multiple LLM platforms with --target

Advanced

• Split 10K+ page docs with router strategy for parallel scraping
• Use three-stream GitHub analysis (code + docs + insights) via Python API
• Configure GitHub profiles for private repos: skill-seekers config --github
• Resume interrupted scrapes: skill-seekers scrape --config config.json --resume
• Custom endpoint support: set ANTHROPIC_BASE_URL for GLM-4.7 compatible APIs

For troubleshooting, see references/advanced-commands.md or references/troubleshooting.md.

Reference Files

Routing Summary

For full skill authoring guidance, load the skill-creator skill.

Technique Map

• Identify scope — Determine what the skill applies to before executing.
• Follow workflow — Use documented steps; avoid ad-hoc shortcuts.
• Verify outputs — Check results match expected contract.
• Handle errors — Graceful degradation when dependencies missing.
• Reference docs — Load references/ when detail needed.
• Preserve state — Don't overwrite user config or artifacts.

Technique Notes

Skill-specific technique rationale. Apply patterns from the skill body. Progressive disclosure: metadata first, body on trigger, references on demand.

Prompt Architect Overlay

Role Definition: Specialist for skill-builder domain. Executes workflows, produces artifacts, routes to related skills when needed.

Input Contract: Context, optional config, artifacts from prior steps. Depends on skill.

Output Contract: Artifacts, status, next-step recommendations. Format per skill.

Edge Cases & Fallbacks: Missing context—ask or infer from workspace. Dependency missing—degrade gracefully; note in output. Ambiguous request—clarify before proceeding.

Agent Rules for Skill Operations

HARD LEARNED (2026-02-11):

1. NEVER use WebFetch to extract SKILL.md content — it summarizes
2. NEVER use haiku agents for skill installation — they hallucinate content
3. ALWAYS git clone source repo and cp files verbatim
4. ALWAYS verify: head -1 == "---", wc -l matches source
5. NEVER add custom frontmatter if source already has it