TBaguette's Atelier

102 skills for Claude Code — judgment calls, code comprehension, change discipline, testing, debugging, systems design, defensive security, communication, and tooling. Every one is project-agnostic, stack-agnostic, and language-agnostic, so the same skill works whether you're in a Rust firmware repo or a Ruby monolith.

Browse them all → lesplooch.github.io/tbaguette-skills

That's the fastest way to see what's actually in here — searchable, organized by category, every skill readable in full.

Install

Claude Desktop, claude.ai, and Cowork

These apps don't install from a folder. They install plugins on your Claude account, from Anthropic's plugin directory: open Customize → Plugins → Discover, in the Desktop app or on claude.ai, and search for TBaguette's Atelier. A plugin added there also turns up in Claude Code at its next session, and the directory keeps it updated.

Until the Atelier's directory listing is live, those apps also take a marketplace you add yourself from a GitHub repository (Anthropic's platform-support table covers where each works), and this repository is one: LeSplooch/tbaguette-skills.

What loads where: the skills load everywhere. The two hooks under What it runs run in Cowork and Claude Code, and chat ignores them, so in chat the skills fire when a task matches rather than being checked before every reply.

Claude Code

macOS, Linux, or Windows in WSL / Git Bash (bash, zsh, and fish all verified — see below):

[ -d ~/.claude/skills/TBaguette/.git ] && git -C ~/.claude/skills/TBaguette pull || git clone https://github.com/LeSplooch/tbaguette-skills.git ~/.claude/skills/TBaguette

Windows, native PowerShell (5.1 or 7+ — the default terminal since Windows 10):

if (Test-Path "$HOME\.claude\skills\TBaguette\.git" -PathType Container) { git -C "$HOME\.claude\skills\TBaguette" pull } else { git clone https://github.com/LeSplooch/tbaguette-skills.git "$HOME\.claude\skills\TBaguette" }

Restart Claude Code (or run /reload-plugins). It loads as the tbaguette-atelier@skills-dir plugin — invoke any skill directly (tbaguette-atelier:formidable, tbaguette-atelier:knowing-when-to-stop, ...) or let them trigger automatically when your situation matches. Run the same command again later to pull updates in place.

New to it? Getting started is the first session in the order you actually hit it: the reload, why the conversation you installed from is the one that won't pick it up, how to prove it landed, six skills worth trying on purpose, and what to check when nothing seems to be happening.

These commands install for Claude Code specifically. The Claude Desktop app, claude.ai and Cowork don't read ~/.claude/skills/ at all, so cloning this repo won't make these skills show up there; they come from the directory, as above.

Both commands only ever touch ~/.claude/skills/TBaguette — neither can alter, merge into, or overwrite any other skill or plugin you already have. git clone refuses outright if that exact path already exists and isn't empty or a clone of this repo, so a name collision fails loudly instead of silently overwriting something. scripts/test_install_command.py (stdlib-only Python, no bash required to run the test) proves this against four scenarios (fresh install, re-run, an empty pre-existing directory, and a real collision), then cross-checks the literal bash command above against every POSIX-ish shell it finds on the build machine — bash, zsh, fish, and sh. All of it is part of run_tests.py, not just asserted here. The PowerShell command isn't machine-tested the same way (no PowerShell runtime in this project's build) — it's verified by careful construction against documented Test-Path/git behavior instead, which the verify-install page is upfront about rather than overclaiming. Read the full walkthrough, the exact test source, and a Command Prompt equivalent too at lesplooch.github.io/tbaguette-skills/verify-install.

Other agents

The commands above are the Claude Code install. The Atelier isn't Claude Code-only, though: this same repo ships an integration for Codex, Cursor, Devin, GitHub Copilot in all three of its forms — the CLI, VS Code, and the coding agent — plus Gemini CLI, Hermes, Kimi Code, OpenCode, and Pi — see PORTING.md for what each one loads and how, plus README.opencode.md and README.kimi.md for the two with enough surface to need their own page.

Those installs differ in kind, not just in path — a plugin command here, an extension install there, a line in a config file somewhere else — so rather than ten recipes, the site's install box carries one prompt you paste into whichever agent you use. It works out which harness it's in and takes the matching route: clone-or-pull into Claude Code's skills directory, hand off to the harness's own install command, show you the config line and ask before touching it, or — when it can't identify a directory its harness actually reads — stop and ask, rather than inventing a path. Only the Claude Code path is one this repo can machine-verify, which is exactly why the prompt is written to discover the rest instead of asserting them.

What's in it

• formidable — design craft for every UI stack: web, native mobile, desktop, terminal, CLI output, game HUD, embedded/e-ink, XR, email, print, voice, dense data. Takes the surface past what was asked, measured against the best shipped work in the same domain rather than against the project's current state.
• Judgment and meta — calibrating confidence, red-teaming your own work, knowing when to stop, and karen-and-the-manager, a persona-forced pass that refuses to be satisfied until it's found everything.
• Reading code, landing changes, testing, debugging, designing systems, defensive security, communicating, environment and tooling — 8 more categories, 76 more skills. Full breakdown in CATALOG.md or, better, on the site above.

What it runs

Nearly all of this is text: skill files the model reads when a task matches. The parts that go further are listed here, so nothing about them is a surprise.

• At session start, hooks/session-start adds the text of using-tbaguette to the conversation. If the plugin folder is a git clone, which is what the install commands above create, it also runs git fetch origin master against this repository and tells the model whether an update is waiting. That fetch is the only network call in the plugin's own code, and it downloads without sending anything. When the plugin folder isn't a git clone, which is the case for a plugin added from Anthropic's directory, the check says so and stops.
• On each prompt, hooks/user-prompt-submit adds a one-sentence reminder to check the skills before replying. On Copilot CLI, which only lets a hook rewrite the prompt, the reminder is prepended to what you typed.
• keeping-tbaguette-current acts on that update check. It fast-forwards a git-clone install only when the clone has no local changes, tells you what changed, and appends a line to tbaguette-update-log.md in your harness's own folder (~/.claude/ on Claude Code).
• tending-tbaguette saves lessons worth contributing to ~/.claude/tbaguette-candidates.md. It forks, pushes, or opens a pull request against this repository only after you say yes to that specific contribution.

Where a harness doesn't run plugin hooks, the first two items don't happen and the skills are just text. Claude chat, on claude.ai and in the Desktop app, is one of those. A plugin added from the directory is updated by the directory, not by keeping-tbaguette-current.

This repo is also the site's source

docs/ is the generated static site (GitHub Pages serves it straight from there). skills/ is the actual plugin content. Edit a skill, then:

python3 scripts/generate.py --base-path /tbaguette-skills

regenerates the whole site — stdlib only, no install step. python3 scripts/run_tests.py runs the full test suite (python3 -m unittest discover on its own misses two of the three test files). Design rationale — the palette, the content schema, why it looks like a bakery — is in superpowers/specs/2026-08-13-tbaguette-showcase-design.md.

No setup step after cloning. python3 scripts/run_tests.py wires the git hooks itself the first time you run it, and says so when it does.

The hooks live in .githooks/. pre-commit regenerates the site before every commit, unconditionally — including a CSS- or skills-only change, which is still "the site was updated." It exists because that got missed manually once: a styles-only commit shipped without regenerating, leaving the header's own "Updated" timestamp pointing at the previous commit instead of itself. pre-push refuses a push to master whose skill graph (the Crumb) no longer matches the skills in the same commit, which catches the routes around the first hook: a commit made with hooks off, a merge, a regeneration from edits nobody staged. python3 scripts/crumb_check.py [REV] [--live] runs the same check by hand, --live against the deployed site.

It needs wiring because git will not read it otherwise: hooks come from .git/hooks by default, and core.hooksPath — the setting that redirects it — is local to each clone and cannot be committed. That is deliberate on git's part, since a repository able to install its own hooks is a repository that runs code on git clone. Nothing can change that; what scripts/githooks.py does is make the wiring happen the first time you run anything here, rather than the first time somebody remembers. It never overwrites a core.hooksPath you set yourself, and to do it by hand instead:

git config core.hooksPath .githooks

Licence

TBaguette's Atelier is free software under the GNU General Public License, version 2. Copyright © 2026 LeSplooch.

That choice is load-bearing rather than decorative. The library is only as good as what it absorbs from real work, so the terms are the ones that keep improvements flowing back instead of accumulating privately in a hundred separate installs. If you change a skill and pass that changed copy on to anyone, GPLv2 requires you to pass on the source under these same terms.

Upstream contribution is a step further than the licence compels, and it is how this repo expects to be used anyway: the tending-tbaguette skill turns a lesson learned while using the Atelier into a pull request here, so it reaches everyone else the next time keeping-tbaguette-current runs.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.