claude-config: Configuração Curatorizada para Claude Code
jajupmochi/claude-config
Baseline reutilizável: regras, skills, hooks e tooling para Claude Code
Ver no GitHubBiblioteca reutilizável com 14 regras de workflow, 7 skills, hooks e recomendações de plugins para padronizar novos projetos Claude Code. Oferece templates de projeto e um skill de inicialização automática (/init-claude-config) que adapta a configuração ao tipo de projeto após responder 6 perguntas.
Instalação
npx convez-ai jajupmochi/claude-config1 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
- Scaffolding automático de novos projetos Claude Code
- Padronizar workflow rules e conventions entre múltiplos projetos
- Compartilhar skills e hooks reutilizáveis com time
Veredito

Config densa e bem estruturada para times que usam Claude Code regularmente. Reduz a fricção do setup inicial, mas exige compreensão de workflow rules e bom entendimento do próprio contexto de projeto para tirar proveito máximo. Investimento upfront alto, retorno em produtividade sustentado.
README
A tradução para o português deste README ainda está na fila.
claude-config
Linlin Jia's curated Claude Code configuration: workflow rules, skills, hooks, plugin recommendations, tooling preferences, and project templates. Install once, then
/init-claude-configin any new project to scaffold it with the relevant subset.
Language: English | 中文
Master TOC
- TL;DR
- What this is
- Quick Start
- Repository structure
- The 14 workflow rules
- The 7 reusable skills
- The 2 hook recipes
- Recommendations (15 curated lists)
- Project templates
- The setup skill
/init-claude-config - For maintainers
- Publishing to external directories
- Build history
- Contributing
- License
TL;DR
Quickest install — npx (no clone, no Claude Code restart needed):
npx github:jajupmochi/claude-config
Then in any project root, open Claude Code and run:
/init-claude-config
Answer 6 questions → project gets scaffolded with the right rules + hooks + skills + tooling for its type.
6 install methods (npx / interactive /plugin / direct /plugin install / local clone / raw URL @imports / copy-paste prompt to CC) — see USAGE.md §0 for all paths.
What this is
Three years of accumulated Claude Code conventions — rules to apply to every new project, hooks worth trusting, skills worth reusing, plugins worth installing — pulled out of half a dozen real research and web projects (liulian-python, swiss-river-network-benchmark, AI_Mur4Cast, jajupmochi.github.io, plus several frontends) and turned into a single composable library.
Four aims:
- One source of truth. Every new project starts from the same baseline. No more copy-paste drift between
CLAUDE.mdfiles. - Per-project picking. A static-site project doesn't need ML training rules. The
/init-claude-configskill asks which context tags apply (research-pkg,ui-project,static-site,ml-research,web-perf, etc.) and only installs the matching subset. - Human-readable. Each rule, hook, and skill explains why it exists, not just what it does. Someone reading without Claude still gets value.
- Agent-executable. Every tooling and install entry includes copy-paste-able commands so an agent can bootstrap a new machine end-to-end without hand-holding.
Quick Start
For the impatient, here's the local-clone path:
# 1. Clone the lib (one time per machine)
git clone https://github.com/jajupmochi/claude-config.git ~/.claude/claude-config
# 2. Open Claude Code in your new project root
cd /path/to/new-project
claude
# 3. Run the scaffold skill
/init-claude-config
The skill will ask:
| Question | Options |
|---|---|
| Project type | Python research / static personal site / frontend / custom |
| Bilingual policy | EN+zh / EN-only / zh-only / decide later |
| Final-output language | Chinese / English / decide later |
| Context tags (multi-select) | always, research-pkg, ui-project, static-site, ml-research, web-perf, image-or-video-work, docs-site, electron-or-desktop |
| Consumption mode | raw URL / local clone / plugin |
| Personal-preference rules | output-brevity / tool-proactivity / no-reread-files |
After answering, your project gets:
CLAUDE.mdwith@importlines for the chosen rules.claude/settings.jsonwith the matching format-on-edit hooks.claude/skills/populated with relevant general + project-type skills- A starter scaffold (
pyproject.toml,index.html, etc.) if you picked a template
For step-by-step walkthroughs of common scenarios, see USAGE.md.
Repository structure
claude-config/
├── README.md / README.zh.md ← you are here
├── USAGE.md / USAGE.zh.md ← step-by-step walkthroughs
├── INVENTORY.md / .zh.md ← master index of catalogued items
├── CLAUDE.md ← rules for editing the lib itself
├── LICENSE ← MIT
├── docs/
│ ├── PHILOSOPHY.md / .zh.md ← the "why" behind the rules
│ ├── CONSUMPTION.md / .zh.md ← three downstream consumption modes
│ └── CONTRIBUTING.md ← how to add new content
├── rules/ ← workflow rules + index
│ └── <rule-name>/RULE.md + snippet.md
├── skills/ ← general skills + index
│ └── general/<skill-name>/SKILL.md
├── hooks/ ← 2 hook recipes + index
│ └── <hook-name>/{README.md, settings.snippet.json}
├── recommendations/ ← recommendation lists + reference tables
│ ├── cc-plugins.md
│ ├── cc-marketplaces-and-skill-bundles.md
│ ├── cli-tools.md
│ ├── js-{ui-and-design, animation-and-3d, build-test-style, state-data}.md
│ ├── web-auditing.md
│ ├── image-video-pdf.md
│ ├── docs-tools.md
│ ├── {ml-research, orchestra-ml-skills}.md
│ └── reference/{apt-packages, vscode-extensions}.md
├── tooling/ ← 3 toolchain preferences + index
│ ├── python-uv-ruff/
│ ├── node-nvm/
│ └── permissions-allowlist/
├── templates/ ← 2 project starters + index
│ ├── research-package-py/
│ └── personal-cite-static/
├── setup/ ← the /init-claude-config skill
│ └── init-claude-config/SKILL.md
├── .claude/
│ ├── skills/ ← meta-skills for editing the lib
│ │ ├── new-rule/, new-skill/, new-hook/, publish/SKILL.md
│ └── (no settings.json — content-only repo)
└── .claude-plugin/
└── plugin.json ← plugin manifest
The 14 workflow rules
Each ships as RULE.md (full content, rationale, examples, exceptions) + snippet.md (drop-in for downstream CLAUDE.md via @import).
| Rule | Scope | When applies |
|---|---|---|
pre-edit-confirmation | universal | List exact targets + 1-line plan + wait for explicit "go" before any Edit / Write |
phased-planning | universal | Tasks touching 3+ files / >5 tool calls / multi-step → numbered phases with per-phase pause |
plugin-preflight | universal | Verify plugin / skill / command is installed AND not deprecated before invoking |
ui-iteration-loop | ui-project | When user provides a visual reference: 8-iteration loop with chrome-devtools screenshots + 4-axis self-critique |
output-brevity | personal | No end-of-batch recap, don't echo tool output, prefer Edit over Write |
tool-proactivity | personal | Installed plugin / skill / MCP fires without asking when matched (with explicit-approval exceptions) |
no-reread-files | personal | Trust your in-session memory of file contents; re-read only on actual change |
chinese-output | personal | Final user-facing output in Chinese; intermediate stays English |
bilingual-docs | optional | NAME.md + NAME.zh.md convention for human-facing docs (consumer opt-in) |
end-of-turn-marker | personal | Every turn ends with [END:FINAL] / [END:WAIT] / [END:NEEDS_USER] on its own line |
always-on-verification | research-pkg | Before any code / test / results claim, invoke code-verifier + research-critic |
autorun-mode | personal | "autorun" / "全力跑" / "think a lot" + scope → higher-autonomy cadence + multi-pass review + branch hygiene |
multi-round-redesign | ui-project | N-round UI redesign protocol with date-stamped outputs + per-round lens discipline + final spec lock |
latex-edit-policy | research-pkg | Editing .tex/.sty/.cls/.bib: hard fixes direct; soft (content) edits comment-don't-delete with % [orig YYYY-MM-DD] inline backup |
The 7 reusable skills
| Skill | Bucket | Trigger | Purpose |
|---|---|---|---|
verify-template | general | /verify | Run CI gates locally (ruff + mypy + pytest); customize per project |
preview-template | general | /preview | Start local dev server (HTTP, Vite, Next.js, MkDocs, Storybook) |
long-running-tasks | general | auto / /long-running-tasks | Decision tree: background subagent vs Monitor vs explicit timeout |
verify-visual | general | auto on UI changes | chrome-devtools MCP screenshot + 4-axis self-critique against reference |
privacy-redact | general | /privacy-redact <file> | Scan file for usernames, absolute paths, tokens, codenames; redact with placeholders |
code-verifier | general | auto / /code-verifier | Three-layer gate before any "tests pass" / "code works" / "results show X" claim — detects FAKE-RUN patterns |
research-critic | general | auto / /research-critic | Six-question audit on every research claim (falsifiability, design, fair comparison, leakage, proportional conclusion, alternatives) |
Plus: the /init-claude-config setup skill (Phase 8 entry point).
The 2 hook recipes
Each ships as README.md (what / why / install / variants) + settings.snippet.json (drop-in JSON).
| Hook | Event | Matcher | Context | Purpose |
|---|---|---|---|---|
ruff-format-on-edit | PostToolUse | Write|Edit | research-pkg / Python | Auto-format *.py with ruff after every Claude edit |
jq-validate-json | PostToolUse | Write|Edit | static-site / JSON-config | Block next tool call if Claude wrote invalid JSON to */locales/*.json or */data/*.json |
Recommendations (15 curated lists)
Each list has agent-executable install commands, context tags, and "why use this" rationale.
| File | Context | Coverage |
|---|---|---|
| cc-plugins.md | always | 37 Claude Code plugins (workflow, integrations, specialized) |
| cc-marketplaces-and-skill-bundles.md | always | 4 third-party marketplaces + 9 skill bundles via npx skills add (GSAP, shadcn, impeccable, Remotion, baoyu, etc.) |
| cli-tools.md | always (selectively) | System CLIs (jq, gh, ripgrep, fd, …) + Python user CLIs (uv, ruff, mkdocs, hf, …) |
| js-ui-and-design.md | ui-project | Lucide, Radix full set, Chakra UI, lenis, d3, visx, recharts, monaco, tanstack/table, shadcn; icon explorers (yesicon.app, svgl.app) |
| js-animation-and-3d.md | 3d-or-animation | motion, gsap, anime.js, lottie-react, tailwindcss-animate, math-curve-loaders; three, R3F, drei, mediapipe; animated icon catalogues (itshover, useanimations); HTML→video (HyperFrames, Remotion); React Native motion |
| js-build-test-style.md | ui-project | vite, next, electron, vitest, playwright, storybook, tailwindcss, prettier |
| js-state-data.md | ui-project | pinia, zustand, swr, vueuse, vue-i18n, vue-router, next-themes |
| web-auditing.md | static-site / web-perf | chrome-devtools MCP (zero-install default), lighthouse CLI, lhci, pa11y, axe-core |
| image-video-pdf.md | image-or-video-work | sharp, svgo, imagemin, ffmpeg (apt), puppeteer |
| docs-tools.md | docs-site | mkdocs + material, ghp-import, latexmk (apt) |
| ml-research.md | ml-research | huggingface_hub[cli], datasets, gpustat, kaleido, selenium; experiment tracking (MLflow, Weights & Biases, ClearML) |
| orchestra-ml-skills.md | ml-research | 21-category ML skill stack incl. 0-autoresearch-skill meta-orchestrator |
| ai-coding-tools.md | optional | Spec-driven scaffolding (OpenSpec) + paper review (paperreview.ai) |
| cluster-hpc.md | optional | SLURM patterns, free-tier rules, rsync conventions for HPC clusters |
| reference-projects.md | optional | Standalone demos / template projects to study (e.g. mykonos-island-voxels — zero-dependency Canvas 2D isometric builder with painterly assets, layered cache rendering, touch-first UI) |
| reference/apt-packages.md | always (lookup) | Apt packages reference table — never auto-install |
| reference/vscode-extensions.md | always (lookup) | VS Code extensions reference table — never auto-install |
Project templates
Minimal-but-complete starters that the setup skill composes with the right rules / hooks / skills.
| Template | Type | Includes |
|---|---|---|
| research-package-py/ | Python research pkg | CLAUDE.template.md + pyproject.template.toml (research extras: torch / data / logging) + .gitignore + .claude/settings.template.json (ruff hook) + project verify skill |
| personal-cite-static/ | Static personal site (HTML/CSS/JS, i18n, bilingual) | CLAUDE.template.md + index.template.html (i18n-aware) + locales/{en,zh}.template.json + .claude/settings.template.json (jq hook) + project preview / verify-visual / i18n-sync skills |
The setup skill /init-claude-config
setup/init-claude-config/SKILL.md is the entry point. Reads project state, asks 6 questions, composes the right subset:
- Detect — empty dir vs existing project; manifest files present?
- Ask — project type, bilingual, final language, context tags, consumption mode, personal preferences
- Compose — copy template, substitute placeholders, generate
CLAUDE.mdwith filtered@importlines, set up.claude/settings.json+.claude/settings.local.json, install relevant skills - Verify —
jq emptyon JSON files,pyproject.tomlparse check,git status - Report — show what was composed + next steps
Idempotent — safe to re-run when adding new context tags later.
For maintainers
To extend claude-config itself, four meta-skills under .claude/skills/:
| Meta-skill | Purpose |
|---|---|
/new-rule | Scaffold a new workflow rule with frontmatter + RULE.md + snippet.md, update INVENTORY |
/new-skill | Scaffold a new skill with proper frontmatter, update INVENTORY |
/new-hook | Scaffold a new hook recipe (8-step construction flow: dedup → pipe-test → wrap → validate → live-proof → cleanup → handoff) |
/publish | Tag a new SemVer version + push + GitHub release with notes from git log |
Plus docs/CONTRIBUTING.md — formal spec for adding rules / skills / hooks / recommendations / tooling / templates, including the inventory-stay-in-sync requirement.
Publishing to external directories
The repo is also being published to community directories so people can discover it. Status of each channel is tracked in PUBLISHING.md:
| Channel | Status |
|---|---|
GitHub Topics (claude-code, claude-skills, claude-config, claude-plugin, scaffold, workflow-rules, ai-coding, developer-tools) | ✅ done — repo is discoverable via GitHub topic pages |
| hesreallyhim/awesome-claude-code PR | 🟡 fork + branch ready — one-click PR creation: see PUBLISHING.md §1 |
| Anthropic plugin marketplace (clau.de/plugin-directory-submission) | ⚪ manual form — content prepared in PUBLISHING.md §2 |
| claudemarketplaces.com / buildwithclaude.com | ⚪ likely auto-aggregated (~24h after topics set) |
npm registry (npx claude-config) | ⚪ requires npm login + npm publish — see PUBLISHING.md §4. Currently usable as npx github:jajupmochi/claude-config |
Build history
claude-config was built in 11 phases on 2026-04-29 (single-day v0.1.0):
| Phase | Focus | Commit |
|---|---|---|
| P1 | Foundation (README, CLAUDE.md, docs, structure) | 1e94686 |
| P1.5 | Discovery scan: local tools inventory → draft docs/DISCOVERY.md (gitignored) | f4cc5eb |
| P2 | 9 workflow rules distilled from ~/.claude/CLAUDE.md and personal-site CLAUDE.local.md | 314e292 |
| P3 | 2 reusable hooks (ruff format-on-edit, jq JSON validate) | 61e5261 |
| P4 | 5 general-purpose skills (verify, preview, long-running-tasks, verify-visual, privacy-redact) | 5ed2b45 |
| P5 | 12 active recommendation lists + 2 reference tables | b328c84 |
| P6 | 3 tooling categories with agent-executable install steps (python-uv-ruff, node-nvm, permissions-allowlist) | f8e2042 |
| P7 | 2 project templates (research-package-py, personal-cite-static) | 5722a88 |
| P8 | /init-claude-config setup skill | b97b8b7 |
| P9 | LICENSE + 4 meta-skills + GitHub publish | 1673f37 + push |
| P10 | Plugin packaging (.claude-plugin/plugin.json) + final docs polish | (this commit) |
Total surface area: 9 rules + 5 + 4 meta skills + 2 hooks + 12 recommendation lists + 3 tooling templates + 2 project templates + 1 setup skill + bilingual docs + 1 plugin manifest.
Contributing
PRs welcome. Open an issue first to align on scope and which category (rule / skill / hook / recommendation / tooling / template) the change belongs to.
See docs/CONTRIBUTING.md for:
- Formal frontmatter conventions
- The "inventory must stay in sync" rule
- Bilingual policy for the lib's own docs
- Conventional Commits style
License
MIT — see LICENSE.