メニュー
Systematic debugging method — feedback loop first, reproduce and minimize, ranked falsifiable hypotheses, targeted instrumentation, regression-test-before-fix. Use whenever diagnosing a bug or a failing finding.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Database engineering standards for PostgreSQL, MySQL, MongoDB, Supabase and Firebase/Firestore — schema/index design, security, and the per-engine gotchas an agent gets wrong. Loaded on top of backend-craft when a ticket touches the database. Use when designing schema, queries, migrations or data-access.
Next.js App Router full-stack standards — server/client boundaries, the four caching layers, server-action security, route handlers, rendering/streaming, metadata, middleware limits, and Next-specific testing. Loaded on top of react-craft (and alongside backend-craft for server code) when the ticket's stack is Next.js. Use when building any Next.js route, action, or API surface.
Headless QA tooling for driving real test tools from the CLI — Postman/Newman (functional + contract), OWASP ZAP + nuclei (security/DAST, the agent-drivable substitute for Burp), and k6/JMeter/autocannon (load). Use when running a comprehensive QA/security/load sweep against a running app.
Full delivery pipeline — takes a feature/project request, plans it into Jira-like markdown tickets, builds each ticket with specialized developer agents, then runs QA, security and code-audit agents, loops a debugger agent over findings until clean, and closes with a final report. Use when the user asks to build a feature or project end-to-end, or invokes /ship.
Express.js (Node) engineering standards — layering, async error handling (Express 4 vs 5), boundary validation, security middleware, graceful shutdown, and testing. Loaded on top of backend-craft when the ticket's stack is Express. Use when building Express APIs.
NestJS engineering standards — feature-module architecture, DI discipline, DTO validation, the guard/interceptor/pipe/filter split, config, TypeORM/Prisma, and testing. Loaded on top of backend-craft when the ticket's stack is NestJS. Use when building NestJS APIs.
| name | debugging-craft |
| description | Systematic debugging method — feedback loop first, reproduce and minimize, ranked falsifiable hypotheses, targeted instrumentation, regression-test-before-fix. Use whenever diagnosing a bug or a failing finding. |
| user-invocable | false |
Six phases, in order. The discipline is the value — skipping phase 1 is how debugging turns into guessing.
Before theorizing, get a deterministic, fast pass/fail signal that goes red on this bug. Ranked options: a failing test at the nearest seam → HTTP script against the dev server → CLI run with fixture input → headless browser drive → minimal repro harness. If you have this loop, you will find the cause; if you don't, you're guessing.
Shrink to the smallest failing case — remove one element at a time (data fields, config, middleware, steps) until removing anything makes it pass. The remainder points at the cause.
Write 3–5 falsifiable hypotheses ranked by likelihood, each with a testable prediction ("if H1, then adding X will show Y"). Test the cheapest-to-check first. Never fix on an untested hypothesis.
Probes mapped to specific predictions — not scattergun logging. Prefer a debugger
or targeted assertions over log spam. Tag every temporary probe (e.g. // DEBUG:)
so cleanup is mechanical.
Write the regression test before the fix — watch it fail on the bug, then pass with the fix. Fix the root cause minimally; check whether the same root cause exists elsewhere in the codebase and fix all instances. Run the surrounding suite for regressions.
Remove all probes and debug artifacts. Record in the ticket/log: symptom → root cause → fix → verification, plus (when it's structural) what would prevent this class of bug.
