workflow
Meta-skill for editing the framework itself — skills, hooks, commands, rules, CLAUDE.md. Use before changing anything under .claude/ or the routing docs, so the change keeps the system's shape (lazy-load, forge markers, hook auto-discovery, harness).
- License 许可
- MIT
- 最近更新
- 2026-11-10
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- workflow
- description
- Meta-skill for editing the framework itself — skills, hooks, commands, rules, CLAUDE.md. Use before changing anything under .claude/ or the routing docs, so the change keeps the system's shape (lazy-load, forge markers, hook auto-discovery, harness).
02 资源文件 6
- 参考文档claude-md.md1.6 KB
- 参考文档command.md1.6 KB
- 参考文档hook.md2.7 KB
- 参考文档rule.md1.3 KB
- 参考文档skill.md2.0 KB
- 参考文档validate.md1.6 KB
03 SKILL.md 正文
Workflow (Self-Edit Meta-Skill)
This is how the framework safely edits ITSELF. Every artifact under .claude/ (skills, hooks, commands, rules) and the CLAUDE.md router follows a shape the runtime depends on. Editing one blindly breaks discovery, enforcement, or copier update. Read the matching operation before you touch it.
Route, don't dump. Load on demand. Enforce, don't suggest. A system artifact earns its tokens — keep every layer lean and let the layer below hold the detail. (
docs/ARCHITECTURE.md§ Design Principles.)
Operations
Only edit-skill has a command wrapper; the rest are sub-operations you reach by reading this skill. Do NOT add new wrappers.
| Operation | When | Sub-file |
|---|---|---|
| Add / edit a skill | New capability or changed workflow knowledge | skill.md |
| Add / edit a command | New or changed slash-command entry point | command.md |
| Add / edit a hook | New/changed guard, stop check, or action | hook.md |
| Edit a rule | Path-scoped domain guidance | rule.md |
| Edit CLAUDE.md | The session router | claude-md.md |
| Validate a change | Before committing any of the above | validate.md |
System map (what the runtime depends on)
| Layer | Location | Contract |
|---|---|---|
| Router | CLAUDE.md (< 200 lines) |
Copier-managed content lives between <!-- forge:start --> / <!-- forge:end -->; edits OUTSIDE the markers survive copier update. Source: CLAUDE.md.jinja. |
| Skills | .claude/skills/**/SKILL.md |
Lazy-loaded: only frontmatter (name, description) enters context at session start; the full body loads on invocation. Router SKILL.md + lean sub-files. |
| Commands | .claude/commands/<domain>/<name>.md |
Auto-discovered by path → slash command. Thin pointer into a skill. |
| Rules | .claude/rules/*.md |
Frontmatter paths: [...] globs; load only when Claude touches a matching file. Project-owned. |
| Hooks | .claude/hooks/ |
Stop stop/checks/[0-9]*.sh + stop/actions/*.sh are auto-discovered by stop/runner.sh (checks run in order, exit ≠0 blocks; actions never block). PreToolUse guards/*.sh are NOT auto-discovered — register them in settings.json. |
| Tests | .claude/hooks/tests/ |
harness.sh drives a hook with no live session; suites are test-*.sh; run-all.sh aggregates them. |
Copier suffix: a template file rendered with answer variables ends in .jinja (stripped on generation); a generic file is copied verbatim. Match the neighbor you are editing.
/workflow:edit-skill
The general self-edit workflow. The /workflow:edit-skill command routes here; despite the name it covers ANY system artifact — skill, hook, command, rule, or CLAUDE.md.
- Identify the artifact and read it in full, plus one sibling as a shape reference. Meta-work rewards full context — read before editing.
- Pick the operation from the table above and read its sub-file.
- Edit, preserving shape — keep the frontmatter contract, the router-then-sub-file split, and the line budgets. Content lives in the lowest layer that can hold it; higher layers only point.
- Validate — run the
validate.mdchecklist. A hook change also needs a harness test (hook.md). - Record it in the state chain (
context/WORKING.md) like any other change.
For skill-specific structure and line budgets, read skill.md.
Add or edit a command
Thin pointers only — a command reads its skill; it does not embed the workflow. Commands are auto-discovered by path, so a new file IS a new slash command; don't add one unless a genuinely new entry point is needed. Full patterns: command.md.
Add or edit a hook
A guard (PreToolUse), a stop check, or a stop action. Stop checks/actions are auto-discovered by stop/runner.sh; guards must be registered in settings.json. EVERY hook change ships with a harness test (test-*.sh) and a green run-all.sh. Full workflow: hook.md.
Edit a rule
Path-scoped guidance that loads only for files matching its paths: globs — cheaper than CLAUDE.md for anything not universal. Full patterns: rule.md.
Edit CLAUDE.md
The router every session reads. Stay under 200 lines; framework content goes between the forge markers (in CLAUDE.md.jinja), project content outside them. Full guidance: claude-md.md.
Validate a change
Before committing any system edit: frontmatter intact, line budgets met, pointers resolve, hooks have passing harness tests, CLAUDE.md under budget. Full checklist: validate.md.
04 同仓库更多 Skills
ponytail
Forces the laziest solution that actually works, simplest, shortest, most minimal. Channels a senior dev who has seen everything: question whether the task needs to exist at all (YAGNI), reach for the standard library before custom code, native platform features before dependencies, one line before fifty. Supports intensity levels: lite, full (default), ultra. Use on ANY coding task: writing, adding, refactoring, fixing, reviewing, or designing code, and choosing libraries or dependencies. Also use whenever the user says "ponytail", "be lazy", "lazy mode", "simplest solution", "minimal solution", "yagni", "do less", or "shortest path", or complains about over-engineering, bloat, boilerplate, or unnecessary dependencies. Do NOT use for non-coding requests (general knowledge, prose, translation, summaries, recipes).
workflow/skill
Patterns for adding or editing a skill — frontmatter contract, router + sub-file split, line budgets.
work
Work lifecycle orchestration: start, continue, pause, handoff, done, plus epic/sprint planning. Context gathering + goal-focused state-chain transitions.
verify
Verification and auditing. Validate acceptance scenarios, run multi-perspective audits, check type-check/build evidence, and gate phase/sprint completion.
team-lead
Coordinate an Agent Team for parallelizable work — implementation, audit, research, or testing. Teaches the mechanical constraints (WAIT pattern, file ownership, model tiers) and lets you design the team. For orchestration_tier=cluster / multi-agent runs.
scope
Plan and design work before building. Invoke when starting an epic, scoping a sprint, or planning any Standard+ change: orient on what exists, equip the planning procedure + domain skills, then produce the SPEC/SCENARIOS/PROGRESS structure and converge it through an audit loop. Feeds /work:epic-plan and the ATDD entry gate.