leek Which Skill
The user does not know which skill fits. Read their situation, then name one skill and say why in a sentence. Recommend, never invoke: most of the skills below are user-invoked, so only the user can start them.
If the situation is genuinely ambiguous, ask one question that separates the branches, not a menu of everything.
The pipeline
Most work runs one straight line (decide → spec → tickets → build) over .scratch/<slug>/. Route by where the user actually is:
| They have… | They want… | Route to |
|---|---|---|
| a vague idea | it stress-tested | grill-with-docs (default in an existing codebase), or grill-me for a plain interview |
| work too big for one session | a map of the decisions | wayfinder |
| an open question needing sources | cited findings | research |
| a question only someone else can answer | a document to send them | to-questionnaire |
| a settled conversation | it written down | to-spec |
| a spec | build ticket files | to-tickets |
| an inbound issue or external PR | it categorised and briefed | triage |
| one ticket | it built end to end | implement |
| a map, spec, or tickets, and no wish to babysit | it driven to done | autopilot |
Before any of it, once per repo: setup writes the tracker, status vocabulary, and domain doc layout.
After the code exists
- Review the diff →
code-review. Want four independent models on it instead of one →panel. - Prove it in the running app →
verify. - Ship the session's work →
commit. - Review comments waiting on a PR →
resolve-review-comments. - Merge or rebase blocked →
resolving-merge-conflicts.
Something is wrong
- Broken, throwing, failing, or slow →
diagnosing-bugs. - The shape of the code is the problem, not one bug →
improve-codebase-architectureto find the opportunity,architecture-satisfactionto run the refactor loop,codebase-designfor the vocabulary either one uses. - Suspect a specific module smells →
code-smells-audit. - Unsure whether a state model or a UI direction holds up →
prototype.
Repo upkeep
housekeeper (low-risk cleanup), repository-cleanup (branches, PRs, stashes, worktrees), dependency-audit (upgrade plan), nightly-docs-sweep (docs vs implementation), distill-sessions (mine session logs for improvements), domain-modeling (terms and ADRs), laravel-herd-worktrees (worktree plus .test URL).
Running out of context
Order matters: the cheapest move that preserves the primary source wins:
- Keep going while the work still fits. Most sessions never need any of the rest.
- Start a fresh session when the next task is genuinely independent. Nothing to carry, nothing to lose.
handoffwhen the next agent needs what this session learned. It writes the document; you start clean with a primary source on disk.- Delegate to a subagent when the expensive part is legwork whose conclusion is all you need back.
- Compact last. Summarising a grilling or diagnosis session replaces a primary source with a confidently-wrong paraphrase. Prefer
handoff, which writes the record down before the context goes.
When nothing fits
Say so plainly and answer the question directly. Reaching for the nearest skill because one must be named is worse than none: it spends a whole workflow on the wrong shape of problem.
Two skills exist for the meta-work rather than the work: writing-for-agents (writing any document an agent consumes) and teach (learn a concept inside this workspace). wait-what is the one to reach for when this answer did not land.