Field Guide · Craft skills

Evaluate without theater

Lightweight eval habits without fake benchmark theater.

The trap

Teams either invent a fake leaderboard or skip evaluation entirely. Both fail. You need a small, honest loop matched to risk.

A lightweight eval

  1. Golden prompt - one saved input that represents a real job
  2. Expected checks - three things that must be true in a good output
  3. Fail examples - one known bad pattern (invented citation, leaked secret tone, wrong owner)
  4. Owner - who says ship or wait
  5. Cadence - when you re-run (model update day, weekly, before launch)

What theater looks like

  • Benchmarks you cannot map to a customer job
  • “The model scored 92%” with no task definition
  • Eval owned by nobody
  • Changing the rubric until the demo passes

On model-update day

Freeze critical workflows long enough to re-run golden prompts. Note diffs. Assign an owner. Silence for a morning beats a broken Friday send. See What should teams do on model-update day?.

Practice

Create one golden prompt for your highest-risk AI workflow. Run it today. Store the output and the pass/fail notes.

See also