fourmi

Spend tokens only where they buy quality. The skill measures first, then changes as little as possible, and never writes a file without the user's go-ahead.

What the user typed after /fourmi: $ARGUMENTS

• Nothing or a path: run steps 1 to 5 on that project (default: the current folder).
• audit: step 1 and step 2 only, change nothing.
• appliquer: steps 1 to 5, still asking before each write.

Answer in the language the user writes in. The blocks exist in French and English (*.md and *.en.md); pick the one that matches the language of the project's existing CLAUDE.md, or of the user if there is none.

1. Measure

python3 "${CLAUDE_SKILL_DIR}/audit_context.py" <project>
python3 "${CLAUDE_SKILL_DIR}/audit_context.py" <project> --json
python3 "${CLAUDE_SKILL_DIR}/audit_context.py" <project> --save /tmp/fourmi-before.json

Always keep the --save snapshot: step 5 compares against it. Use a scratch folder instead of /tmp if the session provides one.

It reports, with token estimates (about 4 characters per token): the CLAUDE.md files loaded on every message (global, project, .claude/CLAUDE.md, CLAUDE.local.md) with their @imports, CLAUDE.md files in subfolders (loaded only when the agent works there), the number of skills and the size of their descriptions, MCP servers (names only, never their settings or secrets), enabled plugins, large files the agent could read by accident, and the permissions.deny rules already in place. It ends with suggestions. It reads files and never changes anything.

2. Explain

Give the user a short reading of the result, in their language:

• the total loaded on every message, and which file weighs most;
• what to trim first, starting with the biggest gain: a long CLAUDE.md (move details into files read on demand or into a skill), MCP servers or plugins this project does not use, skills never used, large files to put behind a Read(...) deny rule;
• what is fine and should stay.

Use the figures from the script; do not invent any. These are estimates, and the real cost also depends on the conversation itself, so say so once. Do not edit ~/.claude.json, settings.json or MCP configuration yourself: show the exact change and let the user decide.

3. Merge the CLAUDE.md block (with the user's go-ahead)

Where it goes: the CLAUDE.md at the project root (<project>/CLAUDE.md). This tunes one project without touching the rules the user applies everywhere; the global ~/.claude/CLAUDE.md is only for habits that hold in every project (step 4 uses it for the chore line). If the project has no CLAUDE.md, the merge creates it.

claude-md-block.md (French) and claude-md-block.en.md (English) hold a short block between <!-- fourmi:start --> and <!-- fourmi:end -->. Both carry the same rules:

• Model routing by difficulty. Haiku: quick research, inventories, summaries, link checks. Sonnet: simple edits, tests, docs, technical SEO, accessibility, commits. Opus: complex implementation, changes over several files, refactors. The tier above Opus (architecture, ambiguous strategy, hard bugs, final review) only if the user asks, since it can be billed as extra credits; otherwise Opus does that work.
• Pass the model parameter on every Agent call.
• Sub-agents only to isolate context or to run truly independent tasks in parallel. Work directly on simple edits, sequential steps and tasks that need one shared context. Plan first; the main agent keeps the plan, the final decisions and the integration. Never run parallel changes on the same files. Re-read source files yourself when a decision depends on them: a sub-agent report summarises isolated work, it does not replace essential material.
• Reading habits and one direction check at the start of big work.

Show the block to the user and adapt it to the project if needed (for example, remove a rule the project's CLAUDE.md already states).

Preview the change first, then write only after a clear yes:

python3 "${CLAUDE_SKILL_DIR}/merge_block.py" <project>/CLAUDE.md "${CLAUDE_SKILL_DIR}/claude-md-block.md"
python3 "${CLAUDE_SKILL_DIR}/merge_block.py" <project>/CLAUDE.md "${CLAUDE_SKILL_DIR}/claude-md-block.md" --write

If the file already has a fourmi block, it is replaced in place; otherwise the block is added at the end. The rest of the file is never touched. If the user changed the block, write the adapted version to a temporary file and merge that one. Run the audit again afterwards and give the new figure.

4. Offer the chore sub-agent (with the user's go-ahead)

agent-ouvriere.md defines a small sub-agent named ouvriere: model: haiku, tools Read, Grep, Glob, Bash, Edit. It handles chores that need no judgement (find and list files, read logs, rename, make a small precise edit, summarise a file) and hands back control as soon as a decision is needed. Explain this in two or three lines, show the file, and ask whether to install it for all sessions.

If yes:

python3 "${CLAUDE_SKILL_DIR}/merge_block.py" ~/.claude/agents/ouvriere.md "${CLAUDE_SKILL_DIR}/agent-ouvriere.md" --copy
python3 "${CLAUDE_SKILL_DIR}/merge_block.py" ~/.claude/agents/ouvriere.md "${CLAUDE_SKILL_DIR}/agent-ouvriere.md" --copy --write

If ~/.claude/agents/ouvriere.md already exists and differs, show the diff and ask before replacing it.

Then offer, as a separate question, to add the "task split" section from delegation-line.md (or delegation-line.en.md) to ~/.claude/CLAUDE.md so the main agent knows when to hand chores over: chores go to ouvriere, while planning, important choices and hard bugs stay with the main model. Same pattern: preview, then --write only after a yes.

python3 "${CLAUDE_SKILL_DIR}/merge_block.py" ~/.claude/CLAUDE.md "${CLAUDE_SKILL_DIR}/delegation-line.md"

A sub-agent defined in a file like this is an official Claude Code feature, not a workaround. Tell the user how to check that it is there:

1. Restart Claude Code if the session was already open (agents are loaded at startup).
2. Type /agents and look for ouvriere in the list of user agents.
3. Ask for a chore ("list the Python files in this project") and watch the agent name appear while it runs.
4. They can also call it explicitly: "use the ouvriere agent to list the Python files".

To remove it, delete ~/.claude/agents/ouvriere.md and the fourmi-agent block from ~/.claude/CLAUDE.md.

Honest limits to tell the user: Haiku is a weaker model, so if an answer from ouvriere looks thin, ask again without the agent (the delegation line says so too). How much each model counts against a subscription is not published; API prices only illustrate the idea (a small model costs a fraction of a large one per token), so do not promise a figure.

5. Verify: measure again

Run the audit again and compare with the snapshot from step 1:

python3 "${CLAUDE_SKILL_DIR}/audit_context.py" <project> --compare /tmp/fourmi-before.json

It prints before -> after for the memory loaded on every message, skill descriptions, MCP servers, plugins, large files and deny rules. Report the figures as they are. The block adds a few hundred tokens to CLAUDE.md on purpose: say so, and explain that the gain comes from routing and delegation (cheaper models for chores, less re-reading), which this static count cannot show. If the number went up a lot, offer to trim. Check also that the project CLAUDE.md contains one fourmi block and nothing was duplicated.

Website projects

For a site (landing page, multi-page site, redesign), offer the extra block claude-md-web.md (or claude-md-web.en.md, markers fourmi-web), merged the same way as the main block and only with a yes:

python3 "${CLAUDE_SKILL_DIR}/merge_block.py" <project>/CLAUDE.md "${CLAUDE_SKILL_DIR}/claude-md-web.md"

The routing it states:

• One direction validation up front (references, page plan, assets, key decisions), not one per section.
• Haiku: asset inventories, reference searches, link and image checks.
• Sonnet: SEO and accessibility.
• Opus: multi-page implementation.
• Top tier, only if the user asks: design system and final review; otherwise Opus.
• Useful sub-agents: asset inventories, reference searches, SEO audits.
• Creative direction stays in the main context and is never delegated.

Rules

• Measure before advising; quote the script's figures.
• Every write (project CLAUDE.md, agent file, global CLAUDE.md) needs its own yes. One approval does not cover the next file.
• Never print secrets: MCP server names only, never their environment or headers.
• Do not delete skills, plugins or MCP servers yourself; give the command and let the user run it.
• Saving tokens never justifies skipping a check that the task needs.
• Never recommend the top model tier by default; name it only as an option the user can ask for.
• State the limits when relevant: figures are estimates, Haiku is weaker, subscription cost per model is not published.