Create PRD (AI Coding Agent)

Generate a PRD formatted as an executable build plan for AI coding agents like Claude Code or Codex. This format combines the standard product context (problem, goals, metrics) with implementation-specific sections that an AI coding tool needs to build the feature with minimal back-and-forth.

The key difference from a standard PRD: every section is written to be machine-actionable. Instead of "we need an API endpoint," this format says "create a POST endpoint at /api/v1/notifications/preferences that accepts this request body and returns this response."

When to Use

• You are handing a feature to Claude Code, Codex, or another AI coding agent to build
• You want a spec detailed enough that the AI can start coding without asking clarifying questions
• You need to bridge the gap between product intent and implementation detail in a single document
• You are a PM who works directly with AI coding tools and wants to give them the best possible instructions

When NOT to Use

• The audience is humans, not AI agents -- use create-prd (standard) or create-prd-engineering
• You need a concise document for leadership approval -- use create-prd-one-pager
• The feature is still in discovery and the solution is not defined enough for implementation detail

Inputs

Ask the user for:

1. Feature name -- What are we building?
2. Problem statement -- What user problem does this solve?
3. Goals and success metrics -- What does success look like?
4. Solution description -- What is the proposed solution? (Must be specific enough to implement.)
5. Codebase context -- What repo, language, framework, and key patterns should the AI follow?
6. Existing code references -- Key files, modules, or patterns the new code should match
7. API contracts -- Endpoints, request/response shapes, error codes (if applicable)
8. Test expectations -- What should be tested? What are the expected outcomes?
9. Constraints -- Performance requirements, backwards compatibility, migration needs

If the user provides partial input, fill in what you can and mark implementation-critical gaps with [BLOCKED -- need X before coding] rather than silent [TBD].

Outputs

A complete PRD + implementation plan in markdown, following the template in references/template.md. The document includes:

• Product context (problem, goals, metrics) -- for the AI to understand intent
• Implementation context (repo, stack, patterns) -- for the AI to match existing code
• File structure -- what files to create or modify
• API contracts with request/response examples -- exact shapes, not descriptions
• Data model changes -- schema additions or modifications
• Build sequence -- ordered steps the AI should follow
• Test cases with expected outcomes -- what to test and what "pass" looks like
• Migration and rollback -- how to deploy safely

Steps

1. Gather product context -- Ask for problem, goals, and success metrics. This grounds the AI in "why" so it makes better implementation decisions.
2. Map the codebase -- Ask what repo, language, and framework. If the user provides a repo path, scan it for patterns (directory structure, naming conventions, test framework).
3. Define the file plan -- List every file that will be created or modified. For modified files, note what changes.
4. Write API contracts -- For each endpoint or interface, write the exact request/response shape as code blocks. Include error cases.
5. Describe data model changes -- Schema changes, new tables/columns, migration scripts needed.
6. Sequence the build -- Order the implementation steps so the AI can work top-to-bottom without circular dependencies. Earlier steps should not depend on later ones.
7. Write test cases -- For each major behavior, write a test case with input, expected output, and the assertion. Include edge cases.
8. Plan migration and rollback -- If the feature requires data migration or has deployment risk, document the steps.
9. Add guardrails -- Note any patterns the AI should follow, anti-patterns to avoid, and things that must NOT change.
10. Review with the user -- Walk through the spec. The key question: "Could an AI coding agent build this feature from this document alone?"

Example Usage

User: "I need a PRD for Claude Code to build a notification preferences API. Users should be able to set per-channel preferences (email, Slack, in-app) for each notification type."

Agent: Asks about the codebase (repo, framework, ORM), existing notification system code, database schema, and auth patterns. Then produces a PRD with exact endpoint specs, database migration SQL, file paths to create, test cases, and a step-by-step build order.

User: "Write a spec for Codex to add SAML SSO to our Express.js app. Here's the repo: ~/code/auth-service."

Agent: Scans the repo for existing auth patterns, middleware structure, and test framework. Produces a PRD that references existing files, matches the code style, and sequences the build so Codex can work through it in order.

Guardrails

• Be explicit, not descriptive. "Create a REST endpoint" is not enough. "Create a POST endpoint at /api/v1/foo that accepts { bar: string, baz: number } and returns { id: string, created_at: string }" is what the AI needs.
• Match existing patterns. The AI should not introduce new patterns when the codebase already has established conventions. Reference specific existing files as examples.
• Sequence matters. If step 3 depends on step 1, say so. AI agents work best when they can go top-to-bottom without backtracking.
• Test cases are not optional. Every behavior should have a test case. The AI needs to know what "done" looks like.
• Mark blockers loudly. Use [BLOCKED] for anything the AI cannot proceed without. Silent [TBD] in an AI-targeted spec leads to hallucinated implementations.
• Do not over-specify style. Specify structure and behavior, not formatting preferences. Let the AI match the existing codebase style.

Related Skills

• create-prd -- Standard PRD for human audiences
• create-prd-engineering -- Detailed engineering spec (for humans, not AI agents)
• create-prd-one-pager -- Leadership one-pager
• create-architecture-diagram -- Visualize the system before specifying the build