Bounded AI-Assisted Engineering Review
Use a deliberately constrained context pack so AI assistance can accelerate review without silently inventing platform facts.
Evidence tier describes my hands-on relationship with the content. Official sources describe platform contracts/documentation. Production observations should not automatically be treated as official limits.
An AI-assisted review that labels evidence, inference and unknowns clearly enough for human validation.
Collect the smallest relevant Atlas/environment context.
Keep client secrets and unnecessary personal data out of the prompt.
Build bounded context
Include only the relevant Atlas nodes and, when needed, a sanitized private environment object.
Require provenance classes
Ask the model to separate documented facts, production observations, inference and gaps.
Validate deterministic outputs
Queries, schemas and automation-facing changes still require tooling/tests rather than trust in prose.
Promote lessons back into the Atlas
If review uncovers a repeatable operational trap, encode it as a Pattern, Runbook or Production Bite.
Turn the Recipe into checkable work.
Progress, notes and history stay in this browser.