​
Agent Skills

CrateStack ships a set of agent skills — instruction files that a coding agent loads before it writes CrateStack code. They live in their own repository, cratestack/cratestack-skills, and install with one command.
This works with Claude Code, Cursor, OpenCode, Codex and the other agents the skills CLI supports. The skills are plain SKILL.md files; nothing executes. Install a subset if you prefer:

​
Why these exist

Most of what an agent needs to know about CrateStack is not inferable from the code in front of it. That a denied model read returns 404 and not 403, that the embedded backend parses policies and deliberately does not enforce them, that @isolation is validated and then discarded, that re-layering DefaultBodyLimit on the generated router does nothing in either direction — these are conventions and decisions, not patterns visible in a file. Without them an agent guesses, and a plausible guess about a framework is how you get code that compiles, passes review, and is wrong.

​
What is covered

Start with cratestack. It picks the right facade for the crate you are in and points at the specific skill for the task.

​
They are versioned, and you should check

CrateStack is pre-1.0. Every public crate shares one version and minor releases break, so a fact that holds for 0.12.0 can be false for the 0.9.x you are running. Every skill therefore opens with the version it was verified against, every release-specific fact is marked (since X.Y.Z), and the repository carries a feature-to-release map — including a standing list of surfaces that are declared but inert, which is the category most likely to cost you an afternoon. Check what you are actually on:

​
How they stay current

A framework pull request that announces a user-facing change has to say what happened to both companions — this site and the skills repo. That is enforced in the framework repo by just verify-parity-declaration. Worth reading precisely: that gate checks the declaration, not the parity. It reads the pull request body and cannot see either companion repository. Keeping the content true is a human job, and a drift audit method is written down in the skills repo’s AGENTS.md.

​
Reporting a wrong skill

A wrong skill is worse than a missing one — a missing skill leaves an agent appropriately unsure, while a wrong one has it writing confident, incorrect code with the tooling vouching for it. If a skill disagrees with the framework, the framework is right, and the disagreement is a bug worth filing.