Use when unfamiliar with a section of code and you need the agent to go up a layer of abstraction — produce a map of relevant modules and callers using CONTEXT.md vocabulary, not implementation detail.

Use when moving issues through a small state machine (bug/enhancement × needs-triage / needs-info / ready-for-agent / ready-for-human / wontfix), prepping briefs that AFK agents can pick up cleanly.

Use when synthesising the current conversation and codebase understanding into a PRD published to the issue tracker — no fresh interview, the context you already have IS the input.

Use when breaking a plan, spec, or PRD into independently-grabbable tracker issues — vertical tracer-bullet slices that cut through every layer end-to-end, each one demoable on its own.

Use when building features or fixing bugs test-first — red-green-refactor in vertical tracer-bullet slices, tests verifying behavior through public interfaces so they survive internal refactors.

Use when building throwaway code that answers a single design question — routes to a runnable terminal app (logic/state) or several switchable UI variations (visual) and is deleted once the question is answered.

Use when looking for deepening opportunities — refactors that turn shallow modules into deep ones, informed by CONTEXT.md domain language and prior ADRs. Aim: testability and AI-navigability.

Use when authoring a new agent skill — covers structure, progressive disclosure, when to add scripts, when to split files, and the trigger-rich description format that controls whether your agent picks the skill at all.

Use when adding Husky pre-commit hooks to a repo — installs husky + lint-staged + Prettier, wires `npx lint-staged`, `npm run typecheck`, and `npm run test` into .husky/pre-commit.

Use when bootstrapping a repo so Matt's engineering skills (to-issues, to-prd, triage, diagnose, tdd, etc.) know the issue tracker, triage labels, and domain doc layout — writes AGENTS.md/CLAUDE.md + docs/agents/.

Use when scaffolding course exercise directories (sections, problems, solutions, explainers) that pass `pnpm ai-hero-cli internal lint` — Matt's AI Hero course tooling convention.

Use when refactoring TypeScript TEST files away from `as` type assertions to @total-typescript/shoehorn — fromPartial() for partial type-safe data, fromAny() for intentionally-wrong inputs.

Use when compacting the current Claude Code conversation into a handoff document a fresh agent can pick up — writes to `mktemp -t handoff-XXXXXX.md`, references existing artefacts instead of duplicating them.

Use when you want a relentless grilling session that stress-tests a plan against the project's domain glossary (CONTEXT.md) and prior ADRs, sharpening terminology and updating docs inline as decisions crystallise.

Use when you want the agent to interview you relentlessly about a plan or design — one question at a time, recommended answer included, walking the decision tree until you've reached shared understanding.

Use when you want Claude Code's Bash tool to refuse dangerous git commands (push, reset --hard, clean -f, branch -D, checkout .) via a PreToolUse hook installed in .claude/settings.json or ~/.claude/settings.json.

Use when editing an article draft — split by headings, model content as a DAG of dependencies, confirm the section order with the author, then tighten each section to ≤240 chars per paragraph.

Use when chasing hard bugs or perf regressions — disciplined reproduce→minimise→hypothesise→instrument→fix→regression-test loop, with feedback loop quality treated as the actual skill.

Use when you want the agent to respond in ultra-compressed caveman-speak — a single SKILL.md that drops filler/articles/pleasantries and saves ~75% of output tokens, no plugin or CLI required.

Use when you want to auto-generate an agent-native Go CLI and MCP server from an API spec, HAR file, or website — with local SQLite mirror, FTS5 search, and compound commands the underlying API can't answer natively.