// the find
chrisbanes/skills
Skills for Kotlin, Jetpack Compose, and Android development
A collection of Agent Skills (structured markdown guides) for Kotlin, Jetpack Compose, and Android work, meant to be loaded into Claude Code, Codex, or OpenCode so the agent routes to a focused playbook instead of guessing. It's for teams already using an AI coding agent daily on an Android/KMP codebase, not for anyone looking for a library or sample app.
The skill set is split narrowly on purpose — compose-state-and-effects and compose-performance are separate skills with a documented migration table from the old, coarser taxonomy, which shows the author is actually maintaining boundaries rather than dumping everything into one mega-prompt. There's a real eval harness (evals/run.py) that runs baseline vs. automatic-routing vs. restraint scenarios per skill and records pass rates, which is more rigor than most prompt collections bother with. Frontmatter is schema-validated in CI against the Agent Skills spec, so the skill files have to conform to a real contract instead of drifting by convention. The author (chrisbanes) has direct Compose/Android runtime experience, so content on recomposition, stability, and Flow modeling is more likely grounded than paraphrased from official docs.
Several of the more interesting workflows (implement-with-subagents, run-github-project) hard-depend on skills pulled from other people's repos (Matt Pocock's, Dimillian's, a third-party ponytail repo) that this package doesn't version or control — if those change behavior, the dependent workflow breaks with no signal from this repo. The eval table showing 100% on almost every automatic/restraint cell, judged by another LLM on a small per-skill case count, reads more like a saturated toy benchmark than evidence the routing actually holds up on messy real repos. The whole thing only pays off inside a specific agent ecosystem (skills CLI, Claude Code, Codex, OpenCode) — there's no way to sanity-check the value by just reading code, you have to run it against real tasks to know if the guidance is any good. The repo is tagged as Python, but the actual content is markdown prompts; the Python is just the eval tooling, so the language label is a poor signal of what you're actually installing.