The author outlines a method to document and stabilize a broken, AI-developed codebase with little or no documentation and heavy tech debt, highlighting duplicated business logic, zombie tables/columns, and implicit downstream dependencies that amplify bugs.

The workflow follows a four-phase pipeline—Analyse, Critique, Fix, Sweep check—with an optional Route facts step, and domain-specific agents (e.g., Orders, Products, Payments) guiding each phase.

In this pipeline, Analyse generates compass data, Critique validates claims without editing, Fix implements changes based on findings, and Sweep check audits for gaps, contradictions, and completeness, while Route facts assigns uncategorized findings to the correct domain compass.

The solution deploys a structured, multi-agent documentation workflow inspired by a major tech firm’s approach to mapping tribal knowledge, using parallel agents to extract actionable context from the codebase through Quick Commands, Key Files, Non-Obvious patterns, and See Also cross-references, answering five core questions per module.

Concrete outcomes include identifying and addressing constraints and gaps in the database and codebase, with the method enabling fixes in a single pass rather than accumulating endless manual constraint notes.

The approach is presented as practical and shovel-ready, not a universal silver bullet, aiming to reduce AI-generated tech debt and bring implicit infrastructure and architecture into explicit, research-driven documentation for future AI agents to follow.