How we judge it

Start with the job to be done, then record success criteria, time, human corrections, and failure boundaries. Feature count is never a proxy for outcome quality.

  • The same input and constraints
  • A reproducible workflow
  • Explicit failure examples

What to watch

The real gain is whether it respects repository constraints, keeps scope tight, and produces verifiable results. Fast output that repeatedly breaks conventions is still negative value.

  • Context setup cost
  • Change-scope control
  • Quality of test evidence

Decision

A useful verdict names who it fits, when not to use it, and the cost and data boundaries to verify before adoption.