Implement features from design docs. Use when (1) user references a design doc from docs/design_docs/, (2) user says "build feature" or "implement feature", (3) user references an existing impl plan from docs/for_ai/plans/. Creates implementation plan from design doc if none exists, then builds.
Bug fixing workflow. Use when (1) user reports a bug, (2) user says "fix bug" or "/fix-bug", (3) debugging failing behavior. Reproduce, fix at the right layer, run tests, open PR, resolve comments, log the task. Add a test only when it captures the bug through observable behavior.
Grade shipped product bets against what their discoveries predicted — pull each registered metric from its source of truth, walk the watch items, verdict every falsification line. Use when (1) user says "/outcome-review", "did it work", "grade our bets", "outcome review", (2) a month has passed since the last review, (3) any outcomes-register row's first check date has passed.
Interview a team member to produce a problem-first proposal — the intake artifact for /sdlc and /adhoc-fullstack-feature. Use when (1) user says "/proposal", "write a proposal", or "pitch", (2) a feature request arrives as a sentence and needs shaping before discovery, (3) someone wants to raise a problem or opportunity for the product pipeline.
Triage and resolve PR review comments — from the /ship-pr review agents (security, design, conventions, bug-hunter, data-migration), from humans, or both. Verify each finding, fix or push back, reply on every thread. Use when user says "/resolve-pr-comments" or asks to resolve PR comments.
Ship a finished slice of work as a PR — open a draft PR, run the review-agent sweep (agents comment inline on the PR), triage every finding, then mark ready. The agents replace the external review bot as the review authority. Use when (1) a workflow skill says "ship with /ship-pr", (2) user asks to open a PR for completed work, (3) work on a branch is built + tested and needs review.
Mine first-party sources for problems nobody raised — public reviews, intake capture forms, analytics funnels, trial/churn data, support threads — and turn them into tier-stamped evidence and candidate problem statements. Use when (1) user says "/signals" or "mine signals", (2) a month has passed since the last sweep, (3) a discovery needs user-side evidence and the requester has none.
Lightweight UX thinking for small features. Use when the direction is relatively clear but benefits from sharpening — clarify intent, challenge assumptions, design quickly. Faster than full discovery, but still thoughtful.