How do you tell whether an AI-coded repository is actually complete?
Do not count files or trust a passing build. Selective Intelligence reconstructs the intended system, follows routes, navigation, services, schemas, permissions, tests, deployment, and live surfaces, then reports the highest state each requirement has really reached.
Signals that this is the problem
- Features exist in source files but are not routed, reachable, or usable.
- README claims, tests, branch state, and the deployed product disagree.
- Interrupted chats or handoffs lost the active objective and repeated completed work.
- Multiple partial implementations compete for the same responsibility.
What Selective Intelligence does
- Recover the governing outcome, current source revision, partial effects, and invalidated evidence.
- Map each requirement through intended, specified, modeled, implemented, wired, reachable, usable, verified, and live.
- Trace routes, consumers, providers, data, authorization, flags, tests, build, and deployment together.
- Repair the causal layer and remove obsolete paths when it is safe to do so.
- Reconcile the plan, repository, release proof, and public behavior at one exact revision.
What proves improvement
- Every included requirement and prohibition has an observable surface and current evidence.
- A local build is not reported as deployed, and an HTTP 200 is not treated as exact release identity.
- The handoff names the verified checkpoint, remaining weaknesses, and next safe action without restarting.
Boundary: A repository audit is not completion by itself. When change authority and tools are available, Selective Intelligence continues through implementation and proportional validation.
Try this exact task
Install the public plugin, then paste this into ChatGPT or Codex with the work you want fixed:
Selective Intelligence: audit this unfinished repository, recover the intended product, reuse what works, finish what is missing or disconnected, and prove the real user flow works.
The exact words Selective Intelligence, an unmistakable request for a named responsibility, or any correction, dissatisfaction, or failure feedback directly activates the canonical skill. The skill never widens permission to publish, spend, delete, deploy, or share.