anombyte93/prd-taskmaster

Zero-config goal-to-tasks engine for Claude Code (the Atlas engine). Graded PRD validation, dependency-ordered task graph, evidence-gated execution.

What it solves

It transforms a high-level goal into production-ready code by eliminating the ambiguity and "hallucinated" completion claims common in AI coding tools. It ensures that software requirements are fully specified without placeholders and that every implementation task is backed by verifiable evidence before being marked as complete.

How it works

The system follows a structured pipeline:

  1. Discovery: An adaptive interview process captures constraints one question at a time.
  2. Generation: It writes a Product Requirements Document (PRD) and scores it against deterministic quality checks. If placeholders like "TBD" are found, the PRD is flagged as needing work.
  3. Task Graphing: The PRD is parsed into a dependency-ordered tasks.json file containing complexity scores and subtasks.
  4. Execution: Tasks are implemented in a loop where each must produce execution evidence. The process only ends when a SHIP_CHECK_OK token is emitted, signifying all gates have passed.

Who it’s for

Developers using Claude Code, Codex, or Gemini who want a more rigorous, PM-led approach to AI-driven development and a way to prove that tasks are actually finished.

Highlights

  • Evidence-Gated Execution: Tasks require proof of completion rather than just a claim from the LLM.
  • Graded PRDs: Automated scoring of specifications to prevent vague requirements or placeholders.
  • ** uma Token Economy**: Routes work to the cheapest capable model, escalating to expensive frontier models only upon failure.
  • Vendor-Neutral State: Stores the PRD and task graph as plain files in the repository, allowing for easy swaps between different AI vendors.
  • Cross-Vendor Fleet: Supports multiple model pools (Claude, Codex, Gemini) to avoid reliance on a single provider.