Why can AI-generated code look correct and still be the wrong product?
Because the AI may have optimized a fluent interpretation, a familiar product pattern, or the code that already exists instead of the person’s real job. Selective Intelligence treats that mismatch as a Step 1 failure: recover the intended outcome again, trace it into behavior, then correct the system rather than defending the implementation.
Signals that this is the problem
- The code runs, but the person says, “That is not what I meant.”
- Tests prove internal consistency while the real user journey is still wrong.
- A correction gets treated as a new feature instead of a repair to the active understanding.
- A dashboard, landing page, workflow, or architecture was chosen because it was familiar.
What Selective Intelligence does
- Preserve the person’s exact outcome, prohibitions, and corrections as the authority.
- Generate at least one plausible wrong reading and compare its real consequences.
- Treat existing code and documentation as evidence, not automatic proof of intent.
- Trace intent to requirement, implementation surface, observed behavior, and current proof.
- Fix the causal mismatch and re-run the end-to-end user path at the exact revision changed.
What proves improvement
- The observable result matches the person’s stated job, not only the technical symptom.
- The counterfactual “could every test pass while the person still says this is wrong?” is false.
- Remaining uncertainty, client limits, and untested states stay visible instead of becoming “done.”
Boundary: Selective Intelligence cannot make an ambiguous irreversible product decision on the person’s behalf. It resolves what evidence can settle, makes reversible progress where safe, and asks one plain question only when competing outcomes would materially differ.
Try this exact task
Install the public plugin, then paste this into ChatGPT or Codex with the work you want fixed:
Selective Intelligence: the AI built the wrong thing. Inspect what exists, recover what I actually meant from this conversation, fix the real cause, and verify the corrected result.
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.