Patterns

Shared labels so cases compound into a map instead of a mood board.

Sycophancy

Agree-first posture. Classic line: "You're absolutely right." Useful as social lubricant. Fatal as engineering method.

Empty confidence

High polish, low proof. Reads senior. Collapses when you ask for a failing test or a source.

Code that lies

Happy path demos, invented APIs, TODO-shaped error handling. The compiler clapped. Production did not.

Template prose

Generic multi-section mush. Swap the product name and the essay still ships. Decisions go missing.

Capacity theater

Limits, staged rollouts, and "demand was hard to predict." Ops reality, product storytelling, same tweet.

Harness mishaps

Model card clean. Claude Code or Agent SDK loop weird. Users still feel "Claude regressed."

Quality regression

Something got worse. Community noticed. Official posts arrive after the screenshots.

Outage and selectability

Short windows where a model vanishes from a menu. Billing bugs invited to the afterparty.

Over-refusal

Safety theater for ordinary asks. Not every "I cannot help with that" is wrong. Some are just lazy fences.

Browse tagged cases on the cases page or the timeline.