Generate PRP

Create a complete Product Requirements Prompt that another agent can execute without the current conversation history.

Phase 0: Discovery Triage (Understanding Lock)

Before research, assess the feature file. If it is clear, scoped, and unambiguous, skip directly to step 1.

Enter Understanding Lock mode if at least 2 of these are true:

• The feature file is short or vague
• Critical aspects are missing (target users, success criteria, non-goals, constraints)
• Multiple plausible interpretations would lead to very different implementations
• Feature touches data, auth, money, or external integrations without specs

Rules in Understanding Lock:

• One question at a time, prefer multiple-choice
• Cover purpose -> users -> success criteria -> non-goals -> constraints
• After at most 5 questions, write a 5-7 bullet Understanding Summary with assumptions and open questions
• Wait for explicit user confirmation before continuing
• Fold the answers back into the feature file or a [feature]-clarified.md sibling

Hard gate: do NOT generate the PRP until the Understanding Summary is confirmed.

Workflow

Step 1: Read the feature file

Read the feature file the user points to. If none is given, default to INITIAL.md.

Step 2: Multi-source research (filter -> rank -> synthesize)

Codex does not have a subagent delegation primitive like Claude, so the executing model does the research itself — but applies the same discipline:

Filter — only read what is plausibly relevant to this feature, not every file in every directory:

• AGENTS.md
• CLAUDE.md if present
• DECISIONS.md (skim ADR titles, read in full only the ones touching the feature area)
• ROADMAP.md (current state only)
• PRPs/ (only PRPs marked as dependencies or recently completed in the same area)
• examples/ or template/examples/ (only the patterns matching the feature's stack)
• Relevant codebase files (grep first, read second)

Rank — prioritize sources by signal-to-noise:

1. Files the feature explicitly references
2. Examples matching the same pattern (api, data, ai-agent, web)
3. Recent git history in affected directories
4. ADRs touching the same architectural area

Synthesize — keep an internal compact brief (5-10 bullets) before writing the PRP. Do not paste raw file content into the PRP — extract the load-bearing fact and cite the path.

Step 3: Verify before citing

Before writing anything into the PRP's must-read or files-to-modify sections:

• Confirm each cited file exists at the path used
• Confirm the pattern claim still matches by reopening the file briefly
• Drop any reference that fails verification

Step 4: Build the PRP

Use PRPs/templates/prp_base.md when available. If this is the template repo, fall back to template/PRPs/templates/prp_base.md.

Produce PRPs/<feature-name>.md with:

• overview and dependencies
• must-read files and docs
• critical gotchas
• files to create and modify
• database changes if any
• ordered tasks with validation commands
• final validation loop
• success criteria

Step 5: Quality gate

Verify the PRP before finishing:

• Referenced files exist
• Every task has validation
• Dependencies are explicit
• Confidence is at least 7/10
• If confidence is below 7, do NOT save — either ask the user clarifying questions or expand research on the specific gaps

Rules

• Make the PRP self-contained.
• Prefer exact file paths over vague descriptions.
• Include shared-document sync in success criteria:
  • ROADMAP.md
  • DECISIONS.md if applicable
  • AGENTS.md
  • CLAUDE.md if present
• Do not start implementation in this skill.