Turning one-line goals into shipped code—with receipts

What it does

prd-taskmaster is a Python engine that plugs into Claude Code (and optionally Codex or Gemini) to turn a vague goal into a shipped feature. It runs an adaptive interview to gather constraints, generates a product-requirements document, scores it against deterministic checks for placeholders and structure, then compiles the requirements into a dependency-ordered task graph. Execution proceeds task by task, and the engine refuses to emit a completion token unless every step produces actual evidence—logs, tests, or files—that passes validation.

The interesting bit

The engine treats spec quality like a school assignment: placeholders and TODOs earn a failing grade that blocks execution. It also treats compute like a budget, routing tasks to the cheapest available model CLI and escalating only when a validator complains. The final SHIP_CHECK_OK token is structurally hard to fake—any non-zero exit status in the evidence chain kills it, and the only bypass is an explicit, audit-logged admin override.

Key highlights

• Deterministic PRD grading that hard-fails on placeholders like TODO or {{...}}, blocking execution until the spec is clean.
• Evidence-gated execution where a single non-zero exit status in any task’s proof blocks the final SHIP_CHECK_OK token.
• Token-economy routing across Claude, Codex, and Gemini CLIs, starting with the cheapest model and escalating only on validator failure.
• Backend-neutral tasks.json keeps your spec and graph as plain repo files, so swapping vendors or canceling Pro doesn’t strand your work.
• Free MIT engine runs entirely offline using existing model CLIs; Pro adds fleet orchestration and bundled MCPs but the core engine is not a demo.

Caveats

• Native Windows is unsupported; the atomic state machine requires POSIX file locking.
• The optional research proxy is not bundled, and the npm install path can emit non-fatal warnings if pip is unavailable for the MCP server’s Python dependencies.
• Fleet orchestration and the browser-verification/secrets-vault MCPs are closed-source Pro add-ons; this repo contains only the solo engine.

Verdict

If you are already using Claude Code and tired of agents declaring victory on half-finished work, this gives you deterministic gates and receipts. If you just want a quick script on native Windows or dislike structured product-management workflows, look elsewhere.