skills

Personal skills that refine how the agent writes code for me. The agent picks them up on its own.

Inspired by mattpocock/skills.

Install

npx skills add krake747/skills

Update with: npx skills update.

The flow

Plan

Plans change in seconds, code in minutes. For a small change, agree the approach first. Keep the plan short enough to read. Write open questions down instead of guessing them.

• plan-first. A 20-line plan before code. It compares approaches with tradeoffs, makes one recommendation, and lists open questions.

Build

Readers judge new code against the code next to it. Match the existing patterns first. Then write the code so the reader mostly sees the normal path. Use clear types and named steps. Keep data unchanged under the reader. Turn bad input away early. Add a new piece only when it does real work.

• scaffold. Match the codebase's patterns when writing new code. Copy the shape, not the code. Say out loud where you break from it and why.
• happy-path. Types first, normal path first. Use clear types. Turn bad input away early with guard clauses. Add no extra layers.

Ship

Write one full report of the change. Then shrink it to a title plus body that still reads well in git log. Ask for one round of notes. Check every claim against the final diff. Then create or update the PR.

• draft-pr. Draft, create, or update a PR with title plus body. It reports context, decisions, and risks, proposes the title plus body, refines once, then ships it.

Prose, anywhere

Reading agent output all day is tiring. This skill guides the agent to simpler terms: first clear structure (verb headings, one action per line, each step ending with how to check it), then a voice that reads like a person wrote it, without breaking any of that.

• humanize. Task docs with verb headings and done lines, plus any prose the agent writes or edits: procedures, runbooks, guides, ADRs, replies, summaries, docs, commit text.

Review

Skills rot as models improve. This skill audits a skill or agent docs against the skill quality checklist and reports findings without changing files.

• skill-review. Review a skill or agent docs for trigger precision, progressive disclosure, rigid recipes, and tentative stops. Report only, never edits. Complements writing-for-agents.

Evaluation

Run the keyless Tier 1 quality checks for all skills with scripts/evaluate-skills.sh. See evals/README.md for setup, scope, and the local Tier 2 command.