Agent skills for a structured engineering workflow. Compatible with agentskills.io.
I wrote these to work more efficiently with AI coding agents. They are opinionated, based on 15+ years of building production software, and encode the workflow I actually follow. They took shape while building Acolyte, where generic prompts did not hold up across sessions; the agent would drift, second-guess itself, or produce noise instead of findings.
npx skills add cniska/skills
Or pick individual skills:
npx skills add cniska/skills -s plan
- explore — Ask questions one at a time until the task is clear. If the code can answer, read the code instead of asking.
- plan — Design a feature through dialogue. The plan emerges from conversation, not isolation.
- tdd — Red-green-refactor. One test, one implementation, one refactor pass. Vertical slices, not horizontal.
- issue — Create a GitHub issue. Check for duplicates, show a draft, get approval before creating.
- review — Run all audits below against the current branch diff in one pass.
- pr — Create a pull request. Runs verification and review before opening.
- style-audit — Naming, control flow, pattern consistency. Enforces local conventions, not generic style guides.
- arch-audit — Boundaries, indirection pressure, contract integrity. If a layer carries no policy, remove it.
- docs-audit — Doc drift, terminology gaps, outdated names after refactors.
- security-audit — Trust boundaries, execution safety, secret exposure. Only reports concrete attack paths.
- test-audit — Coverage gaps, edge cases, test quality. Does not demand 100% coverage.
Every audit skill has the same structure: scope, evidence threshold, workflow, output format, anti-patterns. The evidence threshold is what matters. Findings need concrete code references or plausible failure scenarios. No speculative concerns, no generic style dogma.
The skills compose. /review runs all five audits and deduplicates. /pr gates on /review before creating the pull request. You can run any audit individually if that is all you need.
MIT