AI feature discovery

Claude — Capability probing(animated)

Before designing anything, someone spends real time in the model finding out what it can actually do reliably versus what it just sounds like it can do.

Zoom — Desirability interviews

Separate from whether the model can do it — do people actually want an AI doing this step at all, or would they rather stay in control.

Figma+Cursor — Wizard-of-Oz proto(animated)

The AI response gets faked by a human behind the curtain, so the UX around it can be tested before a single model call is wired up for real.

Notion — Eval & guardrails

A written set of what good and unacceptable model output look like, plus the guardrails that catch the unacceptable kind before a user sees it.

→ flagged

Linear — Ship behind a flag

The feature ships behind a flag to a small cohort first, since an AI feature's failure modes tend to surface with real, messy user input, not test data.