SDD Document Builder

Problem

Many codebases lack a trustworthy SDD: entry points are scattered across HTTP, MQ, scheduled jobs, business flows are only known informally, and documentation often reads like unverifiable conclusions without code locations or counter-evidence. sdd-builder turns documentation drafting into a human-driven, evidence-first workflow, where the AI handles retrieval, synthesis, claims, and validation, while humans confirm the key decisions.

How It Works

Core capability: it links project knowledge, business context, code evidence, SDD drafts, and validation reports into a pausable stage flow:
- Pre-knowledge base: scans rules/, architecture/, and domain-knowledge/, then outputs coding conventions, architectural relationships, and domain concepts with code locations attached.
- Stages one through five: collects business inputs and images, explores the codebase, and builds ENTRY-* and FLOW-* claim pools; it then drafts README.md and automatically validates evidence and claims.
- Final output: removes claim material, keeps an evidence-backed document, and supports human correction plus optional refinement into README.purified.md.

Boundary: it suits repositories that need structured technical documentation and may require iterative refinement. If business context is missing, code evidence is weak, or the expectation is a fully accurate one-shot document, humans still need to supply and verify key inputs.