メニュー
Deliver a PRD through orchestrated task decomposition, implementation agents, and independent QA. Use when every acceptance criterion must be verified before completion.
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
Analyze a list of open questions and suggest up to three answers for each. Use when the user wants options, recommendations, or help resolving multiple unanswered questions.
Explain changes made yesterday as a simple list. Use when user asks what changed yesterday, what was implemented yesterday, or needs a non-technical change summary from recent git work.
Pressure-test, reduce, and rewrite prompts for GPT-5.5. Use when the user asks to grill, audit, critique, harden, improve, optimize, simplify, de-noise, or rewrite a prompt, system prompt, developer instruction, agent playbook, or tool-use workflow specifically for GPT-5.5; especially when it may contain no-op instructions, contradictions, vague rules, missing success criteria, missing output rules, missing validation, unsupported claims, or unclear stop rules.
Build shared understanding before action by interviewing the user about ambiguous implementation, product, planning, or design tasks. Use when goal, scope, constraints, risks, success criteria, done state, or next action are unclear; when user wants a thinking partner; or when user needs a handoff-ready task prompt. Do not write PRDs; use to-prd for product requirements.
Detect sloppy-code signals with repo-wide search metrics and sampled evidence. Use when asked to audit codebase quality, "vibe coding", `isRecord`, broad `Record<string, unknown>` guards, fallback usage, excessive null guards, non-null assertions, unsafe type assertions, `node:*` imports in browser/shared code, swallowed errors, TODO debt, or code smell hotspots before refactoring.
Estimate line additions and deletions from a PRD or spec by reading the requirements and inspecting the target repo. Use when the user asks for changed LOC, additions, deletions, diff size, PR size, or implementation size before building.
| name | prd-delivery |
| description | Deliver a PRD through orchestrated task decomposition, implementation agents, and independent QA. Use when every acceptance criterion must be verified before completion. |
Deliver complete PRD scope. Success requires QA to independently pass every acceptance criterion with concrete evidence. Failed, blocked, or unverified criteria keep delivery active.
Orchestrator extracts scope, non-goals, acceptance criteria, dependencies, risks, and required evidence. It then creates independent tasks suitable for parallel execution.
Each task defines:
Orchestrator uses the Task tool to spawn one implementation agent per independent task. Spawn no more agents than useful.
Implementation agents return their delivery reports. Orchestrator combines the results and resolves conflicts, missing dependencies, and integration gaps.
Orchestrator spawns an independent QA agent to verify the synthesized delivery.
If any criterion fails:
Repeat until the success condition is met.
| Criterion | Status | Evidence |
|---|---|---|
| [criterion] | pass / fail | [command, output, screenshot, file, or behavior proof] |
Include commands, test results, manual checks, and QA conclusion as evidence in the checklist or delivery summary.
