AgentSkillsCN

prd-epic-workflow

在实施前，通过聚焦对话澄清模糊需求。当处理GitHub问题时，该功能将自动运行。适用于需求不明确、功能复杂（超过2天），或涉及跨团队协作的场景。在编码前，可提出两个核心问题——“为什么？”（YAGNI检查）与“更简单吗？”（KISS检查），以确保需求清晰。

SKILL.md
--- frontmatter
name: prd-epic-workflow
description: Run the PRD/EPIC workflow (requirements → simplification → decision-complete EPIC + child TASK breakdown with Execution/Proof).

PRD / EPIC Workflow (AI-only pipeline)

Use this skill whenever starting meaningful work, rewriting an EPIC, or splitting an EPIC into child TASKs.

Hard rules

  • •EPICs are never closed by PRs. PRs close TASKs only.
  • •Every TASK must include a literal line near the top: Epic: #<N> (where #<N> is an EPIC issue).
  • •Every EPIC must include a final TASK titled: Execution: run end-to-end + post evidence.
  • •“Solved” means executed end-to-end with concrete evidence, not “code written”.
  • •Do NOT limit questions. Ask as many rounds as needed to be decision-complete.
  • •Be specific and verifiable.
  • •Log durable learnings immediately on discovery using the repo-local skill agent-learnings, and do a quick end-of-task reminder check.

Base rubric (required)

Use $requirements-clarity as the scoring rubric and structure:

  • •Maintain a clarity score (0–100) and iterate until ≥ 90/100 before finalizing the EPIC.
  • •Track clarification rounds.

Adaptations (project-specific):

  • •Output is GitHub issue markdown only (no docs/prds files).
  • •Question format must be multiple-choice so the owner can answer quickly like: 1a, 2c, 3a (+ note...).

What “incredibly thorough” means here

An EPIC/PRD passes only if it is decision-complete:

  • •A dev agent can implement without asking clarifying questions.
  • •QA can prove pass/fail with commands + expected outputs.
  • •If a script/backfill is involved, the plan includes a safe dry-run, a full run across the defined “ALL records” cohort, and a required evidence comment format.

Workflow

0) Preflight (always do this first)

  1. •Load the issue fresh (do not trust memory):
    • •gh issue view <N> --json title,body,labels,comments,updatedAt
  2. •Confirm the hierarchy:
    • •This PRD/EPIC issue: labels must include type:epic.
  3. •Route required reading (based on the “primary area”):
    • •Always read your project’s docs/rules first.
      • •Examples: README.md, CONTRIBUTING.md, AGENTS.md, docs/.
    • •Then read the area-specific docs for the thing you’re changing (API, workers, data model, UI, etc.).
  4. •Check for “do-not-change” constraints that must be repeated in every TASK.

Question style (required)

Ask one decision at a time using this structure:

  1. •The question (what decision we are making) a) Option A (Recommended) — why recommended b) Option B — tradeoff c) Option C — tradeoff d) Other — owner free-form

(Optional) What changes in the plan based on the answer.

If you cannot offer meaningful alternatives, stop and do more discovery first.

1) PRD interrogation (requirements gathering — ask until locked)

Use QUESTION_BANK.md and do not proceed until every applicable section has explicit answers. If something is unknown, it must be labeled as an assumption with an explicit default.

Read: QUESTION_BANK.md (required).

Minimum required outputs from this phase:

  • •Problem statement (2 sentences):
    • •What the user/system observes
    • •Why it matters (risk/impact)
  • •Goal (WHY + expected behavior)
  • •Definitions & invariants (cohort definitions, timestamps/timezones, “ALL records” definition)
  • •Scope (in / out)
  • •Acceptance criteria (explicit pass/fail bullets; no vague words)
  • •Proof/QA plan (commands + expected outputs)
  • •Execution/Evidence gate (mandatory if any backfill/script/run exists)

Hard anti-vagueness rules

Auto-block if any of these happen:

  • •“works”, “should”, “fix”, “handle”, “robust”, “fast”, “better” without measurable criteria.
  • •“Examples” where a complete list is required (e.g., “these are examples of statuses”).
    • •Force the owner to provide the full list or explicitly approve a default allowlist/denylist.

2) Devil’s advocate + simplification (KISS + YAGNI)

You must propose:

  • •Recommended minimal plan (what we ship first, what we cut)
  • •1–2 alternatives with clear tradeoffs (risk/scope/time)
  • •What can go wrong (failure modes) + mitigations (rate limits, idempotence, timezones, partial runs)

If the recommended plan is not obviously minimal, it fails this step.

2.5) Codebase reuse scan (required)

Before freezing the EPIC approach or decomposing into tasks, scan the repo to find reuse candidates:

  • •Similar workflows/scripts/commands (cron runners, CLI scripts, backfill/migration patterns, common helpers).
  • •Concrete file paths + why they are relevant.
  • •If there are multiple plausible reuse paths, ask the owner to choose.

3) Freeze the PRD (convert to EPIC without creating a new issue)

Rewrite the EPIC body to be decision-complete using the template:

  • •Read: TEMPLATES.md (required).
  • •Update the EPIC issue body so it contains:
    • •Goal
    • •Context/links
    • •Definitions/invariants
    • •Scope (in/out)
    • •Acceptance criteria
    • •Proof/QA plan (commands + expected outputs)
    • •Execution/Evidence gate (mandatory if applicable)
    • •Task breakdown checklist (empty until tasks are created)

Then transition the EPIC state per your repo conventions.

  • •If you use labels/states, keep using them, but do not assume specific label names.
  • •The key requirement is: the EPIC/PRD must be clearly marked as “approved/spec locked” before decomposition.

4) Split into child TASK issues (delivery only happens here)

Create child TASK issues; every one must:

  • •Include Epic: #<EPIC_NUMBER> as the first line of the body.
  • •Be small enough to implement and QA in one pass.

Required next step (same agent)

Once the PRD/EPIC is frozen and the owner says “decompose now”, you must run skill:

  • •epic-decompose

Do not create child TASK issues inside this skill if the owner has not explicitly approved decomposition.

Required child TASK types (most EPICs will have all of these):

  • •Implementation tasks (code changes)
  • •Verification/QA tasks (audits, safety checks, regressions)
  • •Execution: run end-to-end + post evidence (mandatory final gate)

Each TASK issue body must include (use templates in TEMPLATES.md):

  • •Goal
  • •Acceptance criteria
  • •Test plan (commands + expected outputs)
  • •Rollout/execution steps (if any)
  • •Evidence to post before closing (mandatory for Execution/Evidence task)

5) “Solved = executed” enforcement (prevents hanging work)

Rules:

  • •EPIC closes only after the Execution/Evidence TASK is closed and includes the required evidence comment.
  • •If there is a backfill/script, “done” requires:
    • •dry-run (sample)
    • •full run across “ALL records”
    • •evidence: totals, coverage %, mismatch/violations, and 3 real IDs proving correctness

6) Required communication hygiene (no silent decisions)

When the PRD is frozen:

  • •Post a short “spec locked” comment in the EPIC with:
    • •the key decisions
    • •key constraints
    • •the final acceptance criteria list
    • •links to all child TASK issues