Autonomously work one task from an issues/ queue end-to-end — explore, implement (TDD when applicable), run feedback loops, commit, update the issue file. Step 3 of the build pipeline (write-a-prd → prd-to-issues → work-issues). Heavyweight (commits code, modifies the issue queue); explicit invocation only. Headless loop runner: bin/loop.sh. Triggers: /work-issues, "work the issue queue", "run an AFK iteration", "do one issue end-to-end".
79
100%
Does it follow best practices?
Impact
—
No eval scenarios have been run
Passed
No known issues
Do exactly one AFK task end-to-end, then stop. Headless wrappers (bin/loop.sh, bin/once.sh) read this same file as their prompt — keep workflow rules here only.
You need two things in context:
*.md under issues/ (or whatever the project uses). AFK-eligible only — skip anything tagged HITL.Headless mode pre-pastes both into the prompt. Interactive mode (e.g. /work-issues) — gather them yourself:
git log -n 5 --format="%H%n%ad%n%B---" --date=short
ls issues/*.md # then read the relevant onesIf issues/ doesn't exist, ask the user where the queue lives or whether to run prd-to-issues first. Don't invent tasks.
If no AFK tasks remain, output exactly <promise>NO MORE TASKS</promise> and stop. The loop runner uses this string to exit cleanly.
Read before editing. Open the issue, files it references, and adjacent code. Don't guess at structure.
If the task fits TDD (testable behavior change), use the tdd skill. Otherwise: smallest correct change, then verify.
Don't bundle unrelated work. If you find a second problem, append a note to its issue file or open a new one — don't fix it now.
Detect and run the project's loops before committing:
package.json (e.g. npm test, npm run typecheck, npm run lint)pytest, ruff, mypy per pyproject.toml / Makefilego test ./..., go vet ./...cargo test, cargo clippyDon't commit a red build. If a loop fails for reasons unrelated to your change, note it in the commit message and the issue file — don't suppress.
One commit per iteration. Message must include:
Do not add Claude as a co-author.
If issues/NNN-*.rubric.md exists alongside the issue, walk its criteria after the commit and report unmet items in the iteration summary. Do not gate the commit on unmet items — surface, don't block. The rubric is a sharper restatement of acceptance criteria; the human decides whether unmet items are blockers or follow-ups.
issues/done/ (create if missing). Sibling *.rubric.md (if present) moves alongside.reset --hard, push --force, branch deletion) unless the issue explicitly authorizes them.STOP sentinel exists; no need to check yourself.be88d6c
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.