Prd

Turns feature requests into evidence-backed PRDs that live under tasks/ and stay updated as you build. It runs discovery first, inspecting your actual code, tests, and docs before writing anything, then breaks work into phases with concrete checkboxes a junior dev or AI agent can execute. The validation policy is smart: it picks the right mix of unit tests, API checks, browser automation, or manual smoke tests based on what you're actually changing, not some one-size-fits-all recipe. Useful when you need implementation plans that won't instantly go stale or when handing work to agents that need more structure than a chat prompt. Splits into multiple files automatically when plans get big.

GitHub stars: 203
last updated: Apr 20, 2026

View on GitHub

Install

$ npx -y skills add golbin/agent-skills --skill prd --agent claude-code

Installs into .claude/skills of the current project.

Files

PRD

Keep the user's outcome, target design, and remaining work in one document. Planning-only requests end with the PRD; continue implementation when already requested.

Define the target

Read the project evidence needed to identify the problem, affected users, and observable success conditions. Ask about missing decisions that materially affect scope or correctness; record reasonable assumptions for other details.

Choose the simplest design that meets the need, correcting the foundation when necessary. Follow Unix philosophy: focused, composable parts, explicit interfaces, and policy and side effects at boundaries. APIs and internal tools also serve users, including developers and operators.

Include compatibility, recovery, retries, or migration for concrete user, data, contract, or failure risks. Keep temporary mechanisms removable with a retirement condition. Avoid speculative features and abstractions.