Finish the existing job first
A stronger model can remove friction without expanding the brief.
Retrospective field note for 2026-06-01, reconstructed on 16 November 2026. The archive date is not a claim of publication at the time.
There is a recurring temptation with a stronger model: enlarge the task immediately. Add another agent, another stage, another dashboard.
Try removing an existing frustration first. The same correction made to every draft. The context repeated every session. The task that stops at a plan when the request was for a finished file. Those are specific failures with visible outcomes.
Keep the acceptance criteria the same while testing the upgrade. Otherwise the model and the target move together, and nobody can say what improved. Once the original workflow is dependable, expand it where the extra scope serves a real need.
The answer can get shorter too. Better reasoning should help the assistant recognise the constraint sooner and avoid unnecessary explanation. A user does not need to read every alternative the system considered.
Opus 4.8 is another opportunity to test that. The immediate gain should be fewer corrections and fewer interruptions in work already being done. There is plenty of ambition in finishing the original job properly.
Anthropic introduced Claude Opus 4.8 on 28 May. Source.