mattwag05/core
Minimal core skills for the shared agent loop, sized to work on small local models.
Close a work session cleanly: update owning docs, file issues, commit only reviewed paths locally, propose checks for mistakes, and hand off. Use at the end of every session or finished batch of changes. Covers the retired session-closeout and project-housekeeping.
Fix a bug or regression by first reproducing it with a check that can fail, then fixing the root cause. Use when something is broken, throwing, failing, wrong, or slow, or the user says debug.
Settle unclear intent before doing any work, by asking the human numbered questions with recommended answers. Use when the goal or done-check is unclear, the request is ambiguous, or the user says "grill me", "stress-test this plan", or "aim higher" (covers the retired ambition skill).
Write HANDOFF.md so a fresh session can continue the work from files alone, and decide whether to hand off at all. Use when context is getting long, a phase ends, work pauses, or the human asks to hand off or continue later.
Change a live system or service safely: owning record, backup you have proven, approval, reversible experiments, and acceptance by real behavior. Use for any config, restart, upgrade, or state change on a running service or host.
Turn a goal into a plan sized to the work: a short PLAN.md for one session, or a SPEC.md plus ordered, independently checkable issues for multi-session work. Use when starting work bigger than a small fix, or when asked to plan, spec, or break down a task. Covers the retired phased-shipping.
Answer a question from primary sources and write the findings to a file with citations. Use when asked to look something up, check what docs or a spec say, or gather facts before deciding.
Review a change from a fresh start against its original intent and write a findings ledger in REVIEW.md. Use when a consequential change is done, before asking to push or merge, or when the user says review.
Decide whether a change is ready to ask the human to push or merge, and prepare the ask (SHIP.md with verdict, checklist, PR body). Use before asking to push, open a PR, or merge. Covers the PR steps of the retired phased-shipping.
Write or edit an agent skill so a small model can follow it, with evals. Use when creating a new SKILL.md, fixing a skill that misbehaves, or reviewing a skill before it ships.
Build new behavior test-first, or refactor safely, one behavior at a time: write a failing test, see it fail, then make it pass. Use when adding a feature, fixing via tests, refactoring, or the user says tdd or red-green.
Make prose read like a careful human wrote it by removing AI tells, and shape replies so the reader can act. Use when writing or editing any prose a person will read: replies, docs, commit messages, emails, READMEs. Covers the retired unslop and i-have-adhd skills.