Claude Code Meta
grdwng/claude-code-meta
Lifecycle plugin: init → dispatch → self-evolve para projetos Claude
Ver no GitHubSuite de 3 skills para o ciclo de vida completo de projetos Claude Code: inicialização, roteamento de tarefas e auto-evolução. Captura padrões comprovados de 111 sessões de uso real em um plugin portável que todo novo projeto herda desde o dia 1.
Instalação
npx convez-ai grdwng/claude-code-meta1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Bootstrap estruturado de novos projetos Claude Code
- Roteamento automático de tarefas por nível de complexidade (L0-L5)
- Auditoria semanal e auto-melhoria contínua de workflows
Veredito

Excelente ferramenta para equipes que usam Claude Code regularmente e querem sistematizar workflows complexos. Reduz fricção ao onboardar novos projetos com padrões já validados; o sistema de roteamento L0-L5 é robusto e bem documentado.
README
A tradução para o português deste README ainda está na fila.
claude-code-meta
A 3-skill lifecycle suite for Claude Code projects. Distills 01_project's proven workflow (111 sessions of real usage) into a portable plugin covering the full project lifecycle: init → harness → self-evolve.
🎯 Why this exists
01_project built a working workflow over 111 sessions: CodeGraph integration, 4 mandatory skills, spec/plan templates, weekly self-audit. But all of that was locked inside one project. This plugin packages those patterns so any new project inherits them from day 1.
🏗️ Architecture
┌──────────────────────────────────────────────────────────────┐
│ claude-code-meta plugin │
│ │
│ ┌────────────────┐ ┌────────────────┐ ┌────────────────┐ │
│ │ init-project │→ │ dispatch │→ │ self-evolve │ │
│ │ (BIRTH) │ │ (LIVING) │ │ (GROWTH) │ │
│ │ │ │ │ │ │ │
│ │ One-time │ │ Every session │ │ Periodic │ │
│ │ bootstrap │ │ L0–L5 router │ │ audit │ │
│ └────────────────┘ └────────────────┘ └────────────────┘ │
│ ↑ │ │ │
│ │ ↓ ↓ │
│ │ ┌─────────────────────────────┐ │
│ └──────────────│ templates/ │ │
│ │ · ideal-workflow.md │ │
│ │ · spec-template.md │ │
│ │ · plan-template.md │ │
│ │ · audit-skills.sh │ │
│ └─────────────────────────────┘ │
└──────────────────────────────────────────────────────────────┘
📁 File layout
claude-code-meta/
├── README.md ← you are here
├── .claude-plugin/
│ ├── plugin.json ← plugin metadata
│ └── marketplace.json ← marketplace metadata
├── skills/ ← 3 lifecycle skills
│ ├── init-project/SKILL.md ← 8-step bootstrap (Step 8 installs slim core)
│ ├── dispatch/SKILL.md ← L0–L5 router
│ └── self-evolve/SKILL.md ← weekly audit + propose updates
├── templates/ ← 4 templates copied to consumer projects
│ ├── ideal-workflow.md ← baseline (stays in plugin)
│ ├── spec-template.md ← copied to project on init
│ ├── plan-template.md ← copied to project on init
│ └── audit-skills.sh ← copied to project on init
└── harness/ ← self-contained harness layer (new in v0.2.0)
├── CLAUDE.md ← plugin repo's own CLAUDE.md
├── rules/ ← 5 slim rules / 253 lines total
│ ├── task-workflow.md ← L0–L5 routing (80)
│ ├── routing-table.md ← keyword → skill (53)
│ ├── escalation-protocol.md ← mid-flight L2→L3 (35)
│ ├── llm-coding-discipline.md (67)
│ └── bug-fixing-discipline.md (18)
├── hooks/ ← 3 hooks (merged into settings.json on init)
│ ├── codegraph-sync.json ← PostToolUse
│ ├── userpromptsubmit-route.json ← UserPromptSubmit
│ └── quality-discipline-stop.json ← Stop (warn-only, v0.3.1+)
└── memory/MEMORY.md ← plugin's own memory index
Self-contained: cloning this repo + running
init-projectin any new project = full harness restored (5 rules / 253 lines + 3 hooks + 3 skills + 4 templates). No per-machine setup required beyondgit cloneandgh auth login.
🚀 Installation
⚠️ There is no "local install" — copying the plugin to your project root does NOT make the skills discoverable. Claude Code's harness reads
~/.claude/plugins/installed_plugins.jsonat session start. You must complete all 4 steps below.
Step 1: Make the plugin a real git repo
# Put the plugin somewhere OUTSIDE your project (e.g. ~/claude-code-meta-plugin/)
mkdir -p ~/claude-code-meta-plugin
cp -r /path/to/claude-code-meta/. ~/claude-code-meta-plugin/
cd ~/claude-code-meta-plugin
git init -q
git -c user.email="you@local" -c user.name="You" add -A
git -c user.email="you@local" -c user.name="You" commit -q -m "v0.1.0"
SHA=$(git rev-parse HEAD)
echo "Plugin SHA: $SHA" # ← note this, you need it
Step 2: Create a marketplace entry
The plugin source needs a .claude-plugin/marketplace.json (a separate file from plugin.json) listing itself:
{
"name": "claude-code-meta",
"owner": {"name": "You"},
"metadata": {"description": "..."},
"plugins": [{
"name": "claude-code-meta",
"source": "./",
"description": "...",
"version": "0.1.0"
}]
}
Step 3: Verify the manifest format
plugin.jsonhas"skills": ["./skills/"]— array of paths, not skill names- Each
SKILL.mdfrontmatter hasname: <skill-name>— no plugin prefix
Step 4: Register globally
Edit ~/.claude/plugins/known_marketplaces.json:
"gordon-claude-code": {
"source": {"source": "github", "repo": "grdwng/claude-code-meta"},
"installLocation": "/Users/you/.claude/plugins/marketplaces/gordon-claude-code",
"lastUpdated": "<ISO8601>"
}
Edit ~/.claude/plugins/installed_plugins.json:
"plugins": {
"claude-code-meta@gordon-claude-code": [{
"scope": "user",
"installPath": "/Users/you/.claude/plugins/cache/gordon-claude-code/claude-code-meta/0.1.0",
"version": "0.1.0",
"installedAt": "<ISO8601>",
"lastUpdated": "<ISO8601>",
"gitCommitSha": "<real SHA from Step 1>"
}]
}
Copy the plugin files to the cache:
mkdir -p ~/.claude/plugins/cache/gordon-claude-code/claude-code-meta/0.1.0
cp -r ~/claude-code-meta-plugin/{.claude-plugin,skills,templates,README.md} \
~/.claude/plugins/cache/gordon-claude-code/claude-code-meta/0.1.0/
Step 5: Verify in a new session
cd /path/to/your-project
claude # start a NEW session
# In the session: /plugin ← should list claude-code-meta
# Or invoke: claude-code-meta:init-project
Why all this? The harness only reads
installed_plugins.jsonat session start. Anything not registered there is invisible. Seememory/plugin-local-install-doesnt-work.mdfor the full debugging journey (2026-06-07 smoke test).
📖 Usage
Birth (once per project)
User: "Initialize this new project with the claude-code workflow"
Claude (invokes claude-code-meta:init-project):
1. Runs codegraph init -i
2. Creates CLAUDE.md from template
3. Creates memory/MEMORY.md
4. Copies spec/plan templates
5. Copies audit-skills.sh
6. Verifies global PostToolUse hook
7. Verifies slim core (5 rules + 3 hooks)
→ "Project initialized. Ready for development."
Living (every development session)
User: "I want to add feature X"
Claude (claude-code-meta:dispatch, automatic at session start):
→ Calls superpowers:using-superpowers (find right sub-skill)
→ Routes by L-level (L0 bypass / L1–L2 TDD / L3 spec+plan / L4 brainstorm / L5 re-dispatch)
→ Stops for user confirmation at each checkpoint
Growth (weekly + on-demand)
Monday 10:07 AM (cron, set up by init-project):
→ Runs gordon-claude-code:self-evolve
→ bash scripts/audit-skills.sh
→ Compares to docs/superpowers/ideal-workflow.md
→ Proposes rule updates IF deviation persists 2+ cycles (suggested mode)
→ Reports to user
🧭 Skill triggering by L-level (v0.3.1)
Skills trigger by routing-level, not blanket enforcement. Cost asymmetry: L0 questions get a direct answer without invoking the superpowers cascade.
| L | Trigger | Skills (in order) |
|---|---|---|
| L0 | Question word, no imperative | (none — direct answer) |
| L1 | 1 file, no spec | superpowers:test-driven-development → superpowers:verification-before-completion |
| L2 | 1–3 files, no schema/dep | superpowers:test-driven-development → superpowers:verification-before-completion |
| L3 | 3+ files / new dep | superpowers:writing-plans → superpowers:executing-plans → superpowers:test-driven-development → superpowers:verification-before-completion |
| L4 | 5+ files / new arch | superpowers:brainstorming → superpowers:writing-plans → superpowers:executing-plans → superpowers:test-driven-development → code-review |
| L5 | Vague / unclear scope | superpowers:brainstorming → spec draft with [ASSUMED] → re-dispatch |
Routing table: harness/rules/routing-table.md. Dispatcher: skills/dispatch/SKILL.md.
🧬 Lifecycle phases
| Phase | Skill | When | Output |
|---|---|---|---|
| Birth | init-project | Once at project start | Project structure + templates + global state verified |
| Living | dispatch | Every dev session | L0–L5 routing, escalation on L2→L3, slim discipline |
| Growth | self-evolve | Weekly + on-demand | Audit report + deviation analysis + suggested rule updates |
The phases are cyclical: growth feeds back into living (rules update) and informs future births (new projects get latest version).
📊 Origin (2026-06-07)
This plugin was extracted from a self-audit session in 01_project. The audit found:
| Skill | Calls/111 sessions | Verdict |
|---|---|---|
superpowers:writing-plans | 70 | 🟢 healthy |
superpowers:brainstorming | 64 | 🟢 healthy |
superpowers:test-driven-development | 20 | 🟠 underused (0.18/session) |
superpowers:systematic-debugging | 18 | 🟠 underused (0.16/session) |
superpowers:using-superpowers | 3 | 🔴 almost never (0.03/session) |
superpowers:verification-before-completion | 0 | 🔴 never |
The 4 mandatory skill rules, audit script, and templates were extracted and packaged here. The Monday 10:07 cron (3adb96d9) was set up to keep self-evolving.
🔄 Migration from v0.2.0
v0.3.0 is a slim-routing refactor. v0.2.0's workflow-harness orchestrator and 14-rule context are replaced by an L0–L5 router and 5 slim rules.
| Surface | v0.2.0 | v0.3.0 | Δ |
|---|---|---|---|
Rules in ~/.claude/rules/common/ | 14 files / ~724 lines | 5 files / ≤300 lines | −9 files, −~424 lines |
| Hooks | 1 (codegraph-sync) | 2 (+ userpromptsubmit-route) | +1 |
| Per-session dispatcher | workflow-harness skill | dispatch skill | replaced |
| L0 early-bypass | none | question words + no imperative | new |
| Mid-flight escalation | none | L2 → L3 when ≥4 callers | new |
| Templates | spec/plan full | spec/plan with L0-L2 skip markers | lighter |
task-workflow.md | 58 lines, no bypass | ≤120 lines, L0–L5 routing table | rewritten |
ideal-workflow.md | long-form narrative | ≤80 lines, L0–L5 + dispatch reference | rewritten |
| Backward-compat | — | init-project detects & removes v0.2.0 rules | one-shot |
3-step upgrade:
- Pull the plugin — in this repo:
git pull. For external installs: re-fetch the marketplace version. - Re-run
init-projectin each consumer project — it overwrites~/.claude/rules/common/with the 5 slim rules and merges the newuserpromptsubmit-routehook intosettings.json. The old v0.2.0 rules (9 of them) are detected and removed. - Restart Claude Code — so the new
UserPromptSubmithook is loaded. Verify with: ask "What does dispatch do?" — should answer directly in ≤5s without invoking the superpowers cascade.
If a project's ~/.claude/rules/common/ still has the 14 old files after step 2, manually rm them and re-run init-project. The detection is by filename list, not a content heuristic.
📚 Related docs
skills/init-project/SKILL.md— full bootstrap procedureskills/dispatch/SKILL.md— L0–L5 routing + escalation rulesskills/self-evolve/SKILL.md— full audit procedure + suggested-mode rulestemplates/ideal-workflow.md— the baseline self-evolve compares against
External references (in 01_project's memory):
memory/plugin-skill-audit-2026-06-07.md— original audit findingsmemory/periodic-audit-cadence.md— why weeklymemory/superpowers-skill-discipline.md— why using-superpowers must be first
📋 Version
0.3.2 (2026-06-12) — adds 5th Stop hook trigger (warn on multi-step sessions without TaskCreate) + .gitignore for runtime artifacts
0.3.1 (2026-06-12) — quality-discipline enforcement (Stop hook warns when mandatory skills not invoked; mandatory skills use full plugin-prefixed names)
0.3.0 (2026-06-11) — slim-routing refactor (L0–L5, dispatch replaces workflow-harness, 5 rules / ≤300 lines)