Knowledge and source of truth
One of the biggest lessons was that a document is not useful merely because it exists. Early versions of the workflow produced project charters, WBS documents, roadmaps, and large planning files. They looked complete, but ChatGPT and Cursor often did not load them during real implementation. As the product changed, those documents drifted from reality.What went wrong
- agents did not automatically read every planning document
- loading every document would have created too much context
- roadmap and scope changed faster than the documents were maintained
- implementation rationale was often left only in conversations
- later sessions could see what the code did, but not always why it changed
The current model
Durable information is split by purpose. GitHub is normative Code, tests, specs, ADRs, release evidence, and repository engineering metadata live with the repository. AI Work Packet is current coordination state It records only what is needed to resume the active workstream now. Athena is derived searchable knowledge It can preserve rationale, decisions, cross-project explanations, and historical context without becoming the runtime authority. This handbook is explanatory It helps humans remember how the system works but does not override the canonical engineering repository.What belongs where
This prevents the same rule from being copied into several competing sources of truth.
Field finding → durable knowledge
A meaningful finding should leave behind a durable invariant, procedure, decision, or requirement—not only a chat summary.Why minimal context matters
The goal is not to make AI read everything. The goal is to make it read the right thing first:- reading nothing and guessing
- loading every document and drowning the task in context