titouannwtt/claude-tools

How it works ·
What's inside ·
Skills
## What is this?
My personal Claude Code setup, versioned and installed on every machine I work on. It is public because
some of you will want to borrow parts of it, and you should: each piece works on its own.
It is built around three goals:
Spend as few tokens as possible. Each task goes to the cheapest model that can do it, at the right
effort. Scripts do the mechanical work for zero model tokens. Large scopes are read in cheap Haiku batches.
Never trade reliability for it. Every subagent gets a self-contained brief with acceptance criteria.
Reviews and hard calls go to Opus. Every skill script has unit tests. When something fails, the rule is:
improve the brief first, then raise the effort, and only then switch to a bigger model.
Keep a box of reusable skills. Agent skills for Claude Code: transcripts, downloads, blocked web pages, site design audits, SEO for
AI assistants, Mac automation, publish-ready repositories. Take one, take them all.
> [!TIP]
> Everything is written in English, and Claude still answers in your language. To force it, add one line to
> your own ~/.claude/CLAUDE.md, for example: Always reply to the user in French.
## How it works
subagent-gate.py is a PreToolUse hook. Before any subagent starts, it checks the model, the effort,
and the length of the brief. Multi-agent workflows ask you first, unless you set
SUBAGENTGATEWORKFLOW=allow.
CLAUDE.md holds the rules: which agent for which task, a 7-part brief template, and a reflex to ask
"light Haiku research?" before scanning a large, vague scope.
Skills are scripts first. The model launches them, reads the result, and only edits the script if it
breaks.
## What's inside
| Path | What it is |
| kit/CLAUDE.md | Global rules: model × effort matrix, agent choice, brief template, git hygiene |
| kit/agents/ | 10 subagents, each pinned to a model and an effort (table below) |
| kit/hooks/subagent-gate.py | The routing hook that enforces the matrix |
| kit/settings.snippet.json | Model pins, hook registration, no AI attribution in commits |
| skills/ | 10 skills, each with a tested script |
| lib/ | Helpers shared by skills (local settings, proxy fallback) |
| tools/ | Pre-commit checks and a one-command skill publisher |
| Agent | Model · effort | Used for |
| scout | Haiku · low | inventories, grep, extracting facts |
| researcher | Haiku · medium | web and GitHub research with sources |
| worker | Haiku · high | fully specified edits, small fixes, simple scripts |
| implementer | Sonnet · high | features, multi-file refactors, tests |
| fixer | Opus · low | well-specified but delicate execution |
| analyst | Opus · medium | debugging, comparing approaches |
| visual-qa | Opus · medium | screenshots, layout, mobile rendering |
| reviewer | Opus · high | independent review, security, regressions |
| architect | Opus · high | cross-system decisions, stubborn bugs |
| designer | Opus · high | UI, visuals, creative work |
## Skills
| Skill | What it does | Needs | claude.ai |
| light-research | Splits a large scope into ~25k-token batches for Haiku subagents | nothing | no (subagents) |
| skill-forge | Decides if a request deserves a skill, then designs, tests and publishes it | git | no |