← Blog
Week of

Retrospective reconstruction · Published · Weekly archive

Keep the difficult examples

Start with work that exposes a failure, then compare the models.

Field NotesRetrospective

Retrospective field note for 2026-01-05, reconstructed on 16 November 2026. The archive date is not a claim of publication at the time.

A landing page that breaks on mobile. A brief with two requirements that contradict each other. A document where the exception is buried on page seventeen. Keep those examples.

They are a better starting point for choosing a model than another polished demo. Give the candidates the same material and see what actually comes back. Does the layout survive? Does the answer notice the contradiction? Does it find the exception before confidently recommending the wrong thing?

Ten cases would be enough to start. Save the inputs, expected results and corrections. Include the time spent fixing the answer in the comparison. Otherwise a cheap, fast response can look economical while quietly consuming the rest of the afternoon.

Keep the results with the model and settings attached. When the next release arrives, rerun the cases rather than reconstructing what the previous system did from memory.

GPT-5.2 was announced on 11 December 2025. This is carry-forward context, not a release from this week. Source.

All posts Work with me