skill-creator（技能创建器）

Skill Creator

Create, improve, and audit AI agent skills. Every skill follows 14 proven structural patterns.

Scope: Skills only. NOT for creating agents (wagents new agent), building MCP servers (/mcp-creator), or running existing skills. This repo uses raw SKILL.md format committed directly to skills/.

Dispatch

Auto-Detection Heuristic

If no explicit mode keyword is provided:

1. Path ending in SKILL.md or directory under skills/ → Develop (existing)
2. Existing skill name + modification verb (improve, refactor, enhance, update, fix, rewrite, optimize, polish, revise, change) → Develop (existing)
3. --from <source> in arguments → Develop (new, from exemplar)
4. New capability description ("I want to build...", "tool that...", "skill for...") → Develop (new) — derive name, confirm before scaffolding
5. "MCP server", "agent", "run" → refuse gracefully and redirect
6. Ambiguous → ask the user which mode they want

Quick Start

wagents new skill <name>           # Scaffold from template
wagents validate                   # Check all skills
wagents eval validate              # Check eval manifests after eval changes
uv run python skills/skill-creator/scripts/audit.py skills/<name>/  # Score quality
wagents package <name> --dry-run   # Check single-skill portability before packaging

Skill Development

Unified process for creating new skills and improving existing ones. Load references/workflow.md for the full procedure.

Repo-Wide / Multi-Skill Planning

Use plan <name> for an existing-skill refinement plan without editing and plan --all or plan repo for a ranked repo-wide planning pass.

Required planning output:

1. baseline audit summary
2. highest-value findings
3. explicit file targets
4. expected score impact
5. approval gate before any edits

For repo-wide planning, produce a ranked queue plus one standalone refinement plan per promoted skill or skill cluster. Do not edit any skill until the user approves the plan.

Load references/refinement-plan.md when producing the standalone refinement-plan packet.

Audit

Score a skill using deterministic analysis + AI review. Load references/audit-guide.md.

Audit All

Comparative ranking of all repository skills. Load references/audit-guide.md § Audit All.

Dashboard

Render visual creation process monitor or audit quality dashboard. Load references/audit-guide.md § Dashboard.

Auto-detects mode from data: phases field → process monitor; skills array → audit overview.

Gallery (Empty Arguments)

Present skill inventory with scores and available actions. Run uv run python skills/skill-creator/scripts/audit.py --all --format table, display results, offer mode menu.

Package

Package skills into portable ZIP files for Claude Code Desktop import. Load references/packaging-guide.md for ZIP structure, manifest schema, portability checks, and cross-agent compatibility.

wagents package <name> --dry-run  # Check a single skill before emitting a ZIP
wagents package <name>            # Single skill → <name>-v<version>.skill.zip
wagents package --all             # All skills → dist/ with manifest.json
wagents package --all --dry-run   # Check portability without creating ZIPs

Hooks

PreToolUse hooks intercept tool calls during skill execution. The hooks: frontmatter field scopes hooks to this skill only — they activate when the skill is loaded and deactivate when it completes.

Post-edit enforcement for this skill:

• SKILL.md edits trigger uv run wagents validate
• evals/*.json edits trigger uv run wagents eval validate
• hook-bearing skill/settings edits trigger uv run wagents hooks validate
• failures surface to the agent instead of being swallowed

Stop hook enforcement:

• runs uv run python skills/skill-creator/scripts/verify.py stop
• validates dirty skill-definition, eval, and hook surfaces before exit
• exits immediately when hook input has stop_hook_active: true to avoid recursive Stop-hook loops

State Management

Creation progress persists at ~/.{gemini|copilot|codex|claude}/skill-progress/<name>.json. Read/write via scripts/progress.py. Survives session restarts. Use --state-dir to override the default location.

Reference File Index

Read reference files as indicated by the "Read When" column above. Do not rely on memory or prior knowledge of their contents.

Core Principles

Conciseness is respect — The context window is shared. Every line competes with the agent's working memory. Earn every line or delete it.

Progressive disclosure — Frontmatter for discovery (~100 tokens), body for dispatch (<5K tokens), references for deep knowledge (on demand), scripts/templates for execution (never loaded).

Self-exemplar — This skill follows every pattern it teaches. When in doubt, look at how skill-creator applies it.

Critical Rules

1. Run uv run wagents validate before declaring any skill complete
2. Run uv run wagents eval validate after changing evals and before declaring the skill complete
3. Run uv run python skills/skill-creator/scripts/audit.py after every significant SKILL.md change
4. Never create a skill without a dispatch table — it is the routing contract
5. Never create a dispatch table without an empty-args handler — unrouted input is a bug
6. Every reference file must appear in the Reference File Index — orphan refs are invisible
7. Every indexed reference must exist on disk — phantom refs cause agent errors
8. Body must stay under 500 lines (below frontmatter) — move detail to references
9. Description must include "Use when" trigger phrases AND "NOT for" exclusions
10. Names must be kebab-case, 2-64 chars, no consecutive hyphens, no reserved words
11. Scripts use argparse + JSON to stdout — no custom output formats
12. Templates are self-contained HTML with no external dependencies
13. Do NOT call wagents docs generate — delegate to docs-steward
14. Do NOT create agents or MCP servers — refuse gracefully and redirect
15. Improving existing skills requires presenting an improvement plan and getting user approval before implementing changes
16. Audit mode is read-only — never modify the skill being audited
17. Update evals when dispatch behavior or modes change — stale evals are invisible bugs
18. plan <name> and plan --all are read-only planning modes — never edit during planning
19. Repo-wide or multi-skill requests require a ranked plan and standalone refinement-plan output before any implementation begins
20. Stop hooks must include a stop_hook_active guard — recursive hook loops are implementation bugs

Canonical terms (use these exactly throughout):

• Modes: "Develop (new)", "Develop (existing)", "Audit", "Audit All", "Dashboard", "Package", "Gallery"
• Steps (Development): "Understand", "Plan", "Scaffold", "Build", "Validate", "Iterate"
• Grade scale: "A" (90-100), "B" (75-89), "C" (60-74), "D" (40-59), "F" (<40)
• Patterns: "dispatch-table", "reference-file-index", "critical-rules", "canonical-vocabulary", "scope-boundaries", "classification-gating", "scaling-strategy", "state-management", "scripts", "templates", "hooks", "progressive-disclosure", "body-substitutions", "stop-hooks"
• Audit dimensions: "frontmatter", "description", "dispatch-table", "body-structure", "pattern-coverage", "reference-quality", "critical-rules", "script-quality", "portability", "conciseness", "canonical-vocabulary", "evaluation-coverage", "validation-contract"