> Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex. > Cross-reference commands use installed skill names in Codex rather than /octo:* slash commands. > Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. > For host tool equivalents, see skills/blocks/codex-host-adapter.md.

STOP - SKILL ALREADY LOADED

DO NOT call Skill() again. DO NOT load any more skills. Execute directly.

PHASE 0: CLARIFICATION (MANDATORY)

Before writing ANY PRD content, you MUST ask the user these questions:

I need to understand your requirements before creating the PRD.

1. **Target Users**: Who will use this? (developers, end-users, admins, etc.)
2. **Core Problem**: What specific pain point does this solve? Any metrics?
3. **Success Criteria**: How will you measure if this succeeds?
4. **Constraints**: Any technical, budget, or timeline constraints?
5. **Existing Context**: Is this greenfield or integrating with existing systems?

Please answer these (even briefly) so I can create a more targeted PRD.

WAIT for user response before proceeding to Phase 1.

If user says "skip" or provides the feature description inline, extract what you can and note assumptions.

PHASE 1: QUICK RESEARCH (Max 2 searches)

Only search if topic is unfamiliar. Limit to 2 web searches max:

• One for domain/market context
• One for technical patterns (if needed)

Do NOT over-research. 60 seconds max for this phase.

PHASE 2: WRITE PRD

Structure:

1. Executive Summary - Vision + key value prop
2. Problem Statement - Quantified pain points by user segment
3. Goals & Metrics - SMART goals, P0/P1/P2 priority, success metrics table
4. Non-Goals - Explicit boundaries (what we WON'T do)
5. User Personas - 2-3 specific personas with use cases
6. Functional Requirements - FR-001 format with acceptance criteria
7. Implementation Phases - Dependency-ordered, time-boxed
8. Risks & Mitigations - Top 3-5 risks with mitigation strategies

PHASE 2.5: ADVERSARIAL PRD REVIEW (RECOMMENDED)

After drafting the PRD but BEFORE self-scoring, dispatch the draft to a second provider for adversarial review. A single-model PRD has blind spots — cross-provider challenge surfaces wrong assumptions, uncovered scenarios, and contradictory requirements.

Dispatch the PRD draft to a different provider (Codex, Antigravity, or Sonnet as fallback) with this prompt:

> "Challenge this PRD. What assumptions are wrong? What user scenarios are missing? What requirements contradict each other? What will the first user complaint be? What risk does this PRD ignore?"

After receiving the challenge:

• Revise the PRD to address valid challenges
• Note dismissed challenges in the Risks section if they have partial merit
• Add to PRD footer: Adversarial review: applied

Skip with --fast or when user requests speed over thoroughness. See prd.md command for full dispatch syntax.

PHASE 3: SELF-SCORE

Score against 100-point framework:

• AI-Specific Optimization: 25 pts (sequential phases, non-goals, structured format)
• Traditional PRD Core: 25 pts (problem statement, goals, personas, specs)
• Implementation Clarity: 30 pts (FRs with codes, NFRs, architecture, phases)
• Completeness: 20 pts (risks, dependencies, examples, doc quality)

PHASE 4: SAVE

Write to user-specified filename or generate based on feature name.

START WITH PHASE 0 CLARIFICATION QUESTIONS NOW.