Agent Skills

An opinionated workflow for coding agents — my own way of working, current best practices, and lessons from real projects — kept only where evals show it helps.

These skills encode how I build software with agents: scope the task, build against checks, get a second pair of eyes on the diff, prove the UI works, and keep CI green. Each one is run with and without itself; a skill that doesn’t change the outcome gets cut. They install into Claude Code, Cursor, Codex, Gemini CLI, Copilot, Windsurf, OpenCode, and any other Agent Skills-compatible tool.

git clone https://github.com/mthines/agent-skills.git
cd agent-skills && bash scripts/sync-symlinks.sh

What you get

Most skills trigger from plain language — describe what you want. The rest are slash commands you type when you want them.

Measured, not assumed

Every skill here has to earn its place in the context window. Each one is run with and without itself on hard, real-usage cases; skills that changed nothing were deleted or folded into the skill that calls them.

• Over 400 KB less prompt. Seven skills and three agents cut; the reviewer’s prompt went from 202 KB to 60 KB; skill descriptions from 44k to 17k characters.
• aw beats a plain Claude Code run. On three real tickets, each pair blind-judged by two independent reviewers, aw won 10 of 12 judgements — 8 of 8 since its current design. Every win came from a defect its fresh-eyes review caught that the plain run shipped. It costs about 1.5–2× a plain run, so reach for it when a missed bug costs more than that.
• /polish simplify went from 0.73 to 1.00 on planted-defect evals, and the rewritten ux ranked first on 3 of 3 real audits against both the old skill and no skill.

The aw runs used Claude Opus 5.5. Method and every run: PR #241.

Featured: autonomous workflow

Implement a dark-mode toggle independently
Add user authentication end-to-end
Refactor the API client to use retry logic — handle it in isolation

1. Sizes the job. Micro (one file) finishes in one pass; Lite (a small change) writes a short brief and verifies UI changes too; Full gets the loop below.
2. Writes a short brief. At most 60 lines: what done means, what must not change, and one sealed, runnable check per goal.
3. Builds until green. In your session, in its own Git worktree, looping until the checks print CHECKS n/n PASS.
4. Gets a fresh-eyes review. A reviewer that did not write the code reads the diff; every finding is fixed or rejected with evidence.
5. Opens a draft PR with a walkthrough and evidence, then gets CI green. Stuck after five attempts, it stops with BLOCKED: instead of guessing.
6. Proves it works. With .agents/aw/ set up, Lite and Full run /aw-verify on your local stack before the push: it runs the brief’s specs, tries to break the change, and traces each criterion from the web event to the API to the database. Its screenshots and report land on the PR, and a telemetry: criterion must be confirmed before the run counts as done.

Every run prints a paste-ready /goal line to let the session drive itself to done, and records what it learned for the next run. How it works inside: full-tier.md · design notes.

Skills

A / name is a slash command; the rest also trigger from plain language. The skills dispatch five agents of their own: pr-reviewer, branch-reviewer, feature-pr-verifier, aw-executor, and aw-tester.

Install

git clone https://github.com/mthines/agent-skills.git
cd agent-skills && bash scripts/sync-symlinks.sh

Every skill is symlinked live into your tool, so git pull is the whole upgrade. No clone? npx skills add https://github.com/mthines/agent-skills --all --agent claude-code installs a frozen copy.

Forks, other tools, the VS Code extension, and per-project plug-ins: docs/setup.md.

Contributing

PRs welcome. Start with docs/development.md and the conventions in CLAUDE.md.

License

MIT