toolchain
Skill2 starsChanged 20 days agoMITView on GitHub
Use before adding, upgrading or removing a dependency, creating or repairing a venv or node_modules, writing a Dockerfile/CI install step, or running project tooling. Always the lockfile-backed manager (uv for Python, the repo's JS lockfile); never bare pip or the system interpreter.
- Installs packages
In this repository10 other files it gives its agents.
- cards-readmeSkill
- cleanupSkill
- deepsearchSkill
- engineer-mindsetSkill
- grill-meSkill
- investigateSkill
- layoutSkill
- presentSkill
- review-prSkill
- sentry-cliSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 29
Never pip install into the system interpreter, never sudo pip, never pip install --user, never conda activate in a repo that has a uv.lock.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill 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
Toolchain — the env is the lockfile, everywhere
The failure this stops: it works on this Mac, fails in CI, and fails differently in the container, because three environments were built three ways. One lockfile per project, committed, and every environment — local, CI, image, GPU cluster — is built from it.
Adding a dependency at all is the last rung: stdlib → platform feature → something already installed → only then a new package, and the commit message says why.
Python — uv, always
| Intent | Command |
| Add a dependency | uv add <pkg> (--dev for dev-only, --optional dev where dev is an extra) — never hand-edit pyproject.toml |
| Install what CI installs | uv sync --locked --extra dev |
| Run anything in the project env | uv run <cmd> |
| One-off tool, not a project dependency | uvx <tool> |
Check the lock matches pyproject.toml | uv lock --check |
- A bare
uv syncuninstalls pytest in these repos — dev deps here are an extra, not a dependency-group, so--group devis also wrong.uv sync --locked --extra devis what CI runs. uv.lockis committed and must agree withpyproject.toml;uv lock --checkfails when it drifted. Never edit a lockfile by hand.- Never
pip installinto the system interpreter, neversudo pip, neverpip install --user, neverconda activatein a repo that has auv.lock..venv/is gitignored, never committed. - Repos with GPU weights pull gigabytes on sync — for a check that needs no GPU, run the single smoke test rather than syncing everything.
JavaScript — the repo's lockfile names the manager
Detect, do not choose: pnpm-lock.yaml → pnpm · package-lock.json → npm · yarn.lock → yarn.
One lockfile per repo; two lockfiles means two machines resolve two different trees, so delete
the stray one rather than letting both live.
| Intent | npm | pnpm | yarn |
| Install exactly the lock (CI, Docker) | npm ci | pnpm install --frozen-lockfile | yarn install --immutable |
| Add a dependency | npm install <pkg> | pnpm add <pkg> | yarn add <pkg> |
| One-off tool | npx <tool> | pnpm dlx <tool> | yarn dlx <tool> |
Lockfile is committed with the change that caused it. node_modules/ is never committed. No
global installs (npm i -g) for anything a project needs — that is an undeclared dependency.
Keep the installable surface small
A library used by many callers should default to few, light dependencies. Push anything heavy or niche (a specific embedding backend, an optional storage driver) behind an extra so installing the core does not pull it in. A script meant to run out-of-process (a git hook, a background job) that duplicates logic from the main package should stay stdlib-only and never import the package back -- and carry one test asserting the duplicated function still matches the original, so the two cannot silently drift apart.
Container and GPU cluster — the container is the env, the lock still rules
- Base image pinned by digest (
python:3.12-slim@sha256:…), not a moving tag, and never:latest. - Inside the image: copy
pyproject.toml+uv.lock, thenuv sync --locked. No loosepip install Xlines in a Dockerfile — that is a dependency nobody can reproduce. - SkyPilot / SLURM tasks install the same way in
setup:; a task thatpip installs a list of packages inline is how "runs on the cluster, broken locally" happens. See the cluster-job skill, if installed. - The image tag that ran a job goes in that run's
README.md(skilllayout), so the run can be rebuilt.
Repos that predate this rule
Some legacy repos in your org still carry requirements.txt or an environment.yml from before
this convention. Do not convert them as a side effect of an unrelated change. Record the debt once —
./ops problem where the repo has it, otherwise a line in docs/problems/ — and migrate to
pyproject.toml + uv.lock the next time work lands in that repo on purpose.
Check that closes a toolchain change
| Change | Check |
| Added / removed a dependency | uv lock --check (or the JS frozen install) + the repo's own test entry point |
| Dockerfile or CI install step | the build runs, and the installed version matches the lock |
| Anything touching a GPU repo's env | say which environment you verified on — local venv is not the cluster |
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.

