Work Issues

work-issues — one autonomous iteration

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.

Inputs

You need two things in context:

• Issues: every *.md directly under issues/ (or whatever the project uses), EXCLUDING sidecar suffixes (*.rubric.md, *.spec.md, *.research.md). Subdirectories like issues/research/ are not recursed into. AFK-eligible only — skip anything tagged HITL.
• Recent commits: last ~5 commits, so you don't repeat or undo prior work.

Headless mode pre-pastes both into the prompt (the wrappers use list_issue_files from bin/work-issues-lib.sh to honor the exclusion). Interactive mode (e.g. /work-issues) — gather them yourself:

git log -n 5 --format="%H%n%ad%n%B---" --date=short
find issues -maxdepth 1 -type f -name '*.md' \
  -not -name '*.rubric.md' -not -name '*.spec.md' -not -name '*.research.md'
# then read the relevant ones

If issues/ doesn't exist, ask the user where the queue lives or whether to run prd-to-issues first. Don't invent tasks.

Done check

If no AFK tasks remain, output exactly NO MORE TASKS and stop. The loop runner uses this string to exit cleanly.

Task selection — pick one, in priority order

1. Critical bugfixes
2. Dev infrastructure (tests, types, dev scripts) — unblocks feature work
3. Tracer bullets — thin end-to-end slice through every layer of a new feature
4. Polish and quick wins
5. Refactors

Exploration

Read before editing. Open the issue, files it references, and adjacent code. Don't guess at structure.

Implementation

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.

Lane preamble (honor if present)

If a lane preamble has been prepended to this prompt — a leading block that defines an ## Allowlist of literal file paths — you may only edit files that match the allowlist. Never silently expand it. If implementation requires editing a file outside the allowlist, append an ## Allowlist additions requested section to the issue file with one bullet per needed path, each naming the path and a one-sentence reason why the lane needs it, then stop and exit the iteration. The orchestrator surfaces these requests at the next checkpoint. When no lane preamble is present (the default, e.g. direct /work-issues invocation), this section does not apply — proceed as usual.

Feedback loops

Detect and run the project's loops before committing:

• Node: scripts in package.json (e.g. npm test, npm run typecheck, npm run lint)
• Python: pytest, ruff, mypy per pyproject.toml / Makefile
• Go: go test ./..., go vet ./...
• Rust: cargo test, cargo clippy
• Other: whatever the README or CI uses

Don'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.

Commit

One commit per iteration. Message must include:

1. Key decisions and why
2. Files changed (high-level summary)
3. Blockers or notes for next iteration

Do not add Claude as a co-author.

Rubric check (if a sibling rubric exists)

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.

Issue housekeeping

• Complete → move to issues/done/ (create if missing). Sibling *.rubric.md (if present) moves alongside.
• Incomplete → append a dated note: what you did, what remains, blockers.

Hard rules

• One task per iteration.
• No destructive git ops (reset --hard, push --force, branch deletion) unless the issue explicitly authorizes them.
• Unexpected repo state (uncommitted work you didn't make, detached HEAD, in-progress merge/rebase) → stop, write an issue describing it, exit. Don't guess.
• The runner halts before this skill is entered when a STOP sentinel exists; no need to check yourself.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: belchman
• Source: belchman/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.