skills
CLAUDE.md29 starsChanged 28 days agoMITView on GitHub
A collection of AI coding assistant skills ("skills") for French government software development standards. Each skill is a self-contained directory with a SKILL.md (frontmatter + instructions) and optional references/ folder with detailed documentation. Skills are installed by copying them into ~/.claude/skills/ (Claude Code) or ~/.config/opencode/skills/ (OpenCode). This is a documentation-only repo — no build system, no dependencies. All files are Markdown; the only scripts are small Python/shell generators (the skills index and the ANSSI catalogue) and the rgaa scripts (official…
In this repository12 other files it gives its agents.
- albert-apiSkill
- anssi-guidesSkill
- datagouv-apisSkill
- presentation-docsSkill
- rag-indexSkill
- rag-parseSkill
- rag-searchSkill
- lasuite-ui-kitSkill
- react-dsfrSkill
- rgaaSkill
- securite-developpementSkill
- usage-ia-agents-etatSkill
The file
We split this CLAUDE.md into the rules it gives an agent
5 rules. About 0 are common, 3 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 50
All content is in French
- Very specific: line 54
After any change to a skill, run python3 scripts/build_index.py to regenerate skills/index.json and skills/.experimental/index.json …
- Very specific: line 58
Commit messages follow conventional commits format (e.g., feat(skill-name):, docs(readme):, refactor(skill-name):)
- Distinctive: line 62
Before considering a task complete, run the tests and verify that they pass
- Distinctive: line 63
If you modify existing code, add tests if there aren't any already
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
What This Repo Is
A collection of AI coding assistant skills ("skills") for French government software development standards. Each skill is a self-contained directory with a SKILL.md (frontmatter + instructions) and optional references/ folder with detailed documentation. Skills are installed by copying them into ~/.claude/skills/ (Claude Code) or ~/.config/opencode/skills/ (OpenCode).
This is a documentation-only repo — no build system, no dependencies. All files are Markdown; the only scripts are small Python/shell generators (the skills index and the ANSSI catalogue) and the rgaa scripts (official text generator, classement, DSFR class and form field name checkers), each with its tests. None of the rgaa scripts runs in CI: run them by hand after editing a reference.
Architecture
Each skill follows the same structure:
skill-name/
├── SKILL.md # Frontmatter (name, description, trigger conditions) + main instructions
└── references/ # Optional detailed reference docs split by topicThe SKILL.md frontmatter fields (name, description) are used by AI tools to decide when to activate the skill. The description field is critical — it defines the trigger conditions. Some skills (rag-*) use additional fields provider and available-providers to support plug-and-play of different backend providers.
Current Skills
Skills are split into two tiers:
Official skills (skills/) — backed by State standards, maintained by the DINUM IAE department:
- react-dsfr — React components for the Design System de l'État (DSFR) using
@codegouvfr/react-dsfr - lasuite-ui-kit — React components for LaSuite apps using
@gouvfr-lasuite/ui-kitand@gouvfr-lasuite/cunningham-react - rgaa — AI-assisted RGAA 4.1.2 self-assessment (106 criteria; produces a structured report that supports an audit, not a conformity audit). Code examples in
references/are copied as-is by readers:skills/rgaa/scripts/check-dsfr-classes.pychecks everyfr-*class in their code blocks against the CSS of the latest published@gouvfr/dsfr(resolved on each run). Run it after editing a reference; its offline test isskills/rgaa/scripts/test-check-dsfr-classes.sh. Likewise,skills/rgaa/scripts/check-form-names.pyflags every<input>,<select>or<textarea>without a non-emptynameinhtml,erbandhtml+erbblocks (a field the example would not submit); a field deliberately not submitted is declared by a<%# sans name : reason %>or<!-- sans name : reason -->comment right before it. Its offline test isskills/rgaa/scripts/test-check-form-names.sh. - securite-developpement — Security reference for government app development, generated by DINUM from 13 ANSSI guides. 14 core domains plus 3 conditional modules (C, Rust, CMS) loaded only when the project matches. Two modes: advice (while coding) and audit (conformity report). Every rule carries its source — an exact ANSSI recommendation id (
[TLS R3]), an unnumbered "Essentiel" ([ESS-BDD]), or[DINUM]when no ANSSI guide covers it. Seeskills/securite-developpement/references/sources.mdfor guide versions and known divergences between guides. - anssi-guides — Finder for the ANSSI catalog (128 French guides plus 1 English-only publication as of 2026-09-29), backed by the MesServicesCyber JSON API: locates the relevant guide through its title, description, collections, needs and theme, fetches document URLs from the API, and answers with exact citations. It locates and quotes, it does not pre-digest rules — the traced security rulebook remains
securite-developpement, and the 13 guide pages covered there are starred inskills/anssi-guides/references/catalogue.md. The checked-in catalog is generated byskills/anssi-guides/scripts/generate-catalogue.sh; a revision of a starred guide must be propagated tosecurite-developpement/references/sources.md. - usage-ia-agents-etat — AI usage framework for State public servants, based on the DINUM Guide d'usage de l'IA pour les agents publics de l'État (5 core principles). Acts as a reference AND proactively warns when the user is about to breach the guide (sensitive data in a consumer AI tool, missing AI-usage disclosure, unattended automation, high-risk RIA use cases)
- datagouv-apis — data.gouv.fr APIs reference (catalog, metrics, tabular) — synced from datagouv/datagouv-skill
Experimental skills (skills/.experimental/) ⚠️ — not maintained by the DINUM, depend on unaudited third-party tools, no continuity guarantee. The .experimental/ subdirectory is a convention recognized by the Vercel Skills CLI so they remain installable. See skills/.experimental/README.md:
- rag-parse — Convert PDF/DOCX/PPTX/XLSX/images to markdown or JSON with LiteParse (local, no cloud dependencies)
- rag-index — Index markdown documents for semantic search with qmd (local SQLite + sqlite-vec)
- rag-search — Search indexed documents with qmd (semantic and keyword search)
Templates
The templates/instructions/ directory contains ready-to-use INSTRUCTIONS.md files for government projects. Each template covers language conventions, design system, accessibility, security, testing, Git conventions, and links to relevant skills from this repo. Two variants exist: beta.gouv.md (beta.gouv ecosystem, DSFR, pnpm) and LaSuite.md (La Suite numérique, UI Kit + Cunningham, yarn). Templates are written in English (instruction language) while preserving French convention values.
Upstream Sync
The datagouv-apis skill is synced from an external repo (datagouv/datagouv-skill). A GitHub Actions workflow (.github/workflows/sync-datagouv.yml) runs weekly (Mondays at 09:00 UTC) and on manual trigger. Only SKILL.md is synced — the skill's README.md is DINUM-authored and intentionally diverges from upstream. Accordingly, .sync-state/datagouv.sha stores the git blob hash of the upstream SKILL.md, not the upstream HEAD commit SHA: the workflow opens a PR only when that file's content actually changes, so upstream commits touching only README/LICENSE do not produce empty sync PRs. This mechanism is extensible to other third-party skills.
Conventions
- All content is in French
- Skills target React/TypeScript projects, except
rgaa, which audits HTML, React/JSX and Rails/ERB code - SKILL.md files use YAML frontmatter (
---delimiters) withnameanddescriptionfields; RAG skills additionally useproviderandavailable-providersfields - Reference files are split by theme/topic to keep individual files manageable
- After any change to a skill, run
python3 scripts/build_index.pyto regenerateskills/index.jsonandskills/.experimental/index.json(checked in CI) - The rgaa skill is an audit tool (input: code to audit → output: structured report), not a code generation skill
- In the rgaa skill, each criterion is described twice: its official text and DINUM guidance in
references/criteres-*.md, its evaluation mode (code, mixte, conditions réelles, jugement humain) inreferences/classement/. Any change to a criterion touches both files; then runskills/rgaa/scripts/check-classement.py(offline test:skills/rgaa/scripts/test-check-classement.sh) - The « Texte officiel » blocks in
skills/rgaa/references/criteres-*.mdandreferences/glossaire.mdare generated fromskills/rgaa/sources/(official DINUM JSON, never edited): never reformulate them by hand. After changing the source or the classement, runpython3 skills/rgaa/scripts/officiel.py --write, thenofficiel.py(check) andsh skills/rgaa/scripts/test-officiel.sh. Per-stack corrections live inreferences/corrections/, same file names - Commit messages follow conventional commits format (e.g.,
feat(skill-name):,docs(readme):,refactor(skill-name):)
Tests
- Any new feature must be accompanied by tests
- Before considering a task complete, run the tests and verify that they pass
- If you modify existing code, add tests if there aren't any already
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

