AI Evals for Designers: Quick Checklist

Five steps to find where your AI product goes wrong, using 20 real answers and a spreadsheet. About two hours, no code.

  1. Step 1

    Collect 20 real answers

    Pull 20 real conversations or outputs from your product. Include the awkward ones: requests it cannot handle, missing information, things it should refuse.

    About 20 min for 20 answers

    For example

    • Support botReal tickets, including refund requests and an angry customer
    • Writing assistantDrafts from very short prompts and off-brand requests
    • Search or summaryQuestions with no good answer in the sources

  2. Step 2

    Mark each one pass or fail

    Ask one question: would I be happy if a real user got this? When it fails, write one plain line on the first thing wrong. No scores and no categories yet.

    About 40 min for 20 answers

    For example

    • Support botFail: promises a refund it cannot issue
    • Writing assistantFail: tone far too casual for a legal email
    • Search or summaryFail: summary leaves out the main caveat

  3. Step 3

    Group the notes and count

    Sort the notes into piles the way you would affinity-map research. Count each pile. The biggest pile is your first problem.

    About 20 min for 20 answers

    For example

    • Support botClaims actions it cannot take
    • Writing assistantIgnores the requested tone
    • Search or summaryDrops important caveats

  4. Step 4

    Turn the biggest pile into one rule

    Write it as a yes or no question that anyone on the team could apply to an answer and get the same verdict.

    About 10 min for 20 answers

    For example

    • Support botDoes it only promise actions the product can actually take?
    • Writing assistantDoes the draft match the tone the user asked for?
    • Search or summaryDoes the summary keep every caveat from the source?

  5. Step 5

    Fix, then re-check the same 20 plus 5 new ones

    Change one thing, rerun the same answers and apply the same rule. Add a few the fix never saw, so you know it works beyond the ones you tuned it on.

    About 30 min for 20 answers

    For example

    • Any productCount how many answers break the rule, before and after
    • Any productRead the new ones too: a fix can quietly break something else

How we did it

The same five steps on our own AI tool

Our audit tool takes a screenshot of an AI product and returns design advice, so it fails in ways a chat or support product would not. Your piles and your rule will be different. What carries over is the method.

  1. 01CollectOur product turns a screenshot into design advice, so our answers were audits: 41 screenshots, including pricing, login and legal pages where there is nothing AI to find.
  2. 02MarkImran marked all 41 audits. An AI gave a second opinion, and he made the final call where they disagreed.
  3. 03GroupSix piles. Two tied for biggest: advice about things a single screenshot cannot show, and asking for confidence scores on ordinary chat.
  4. 04Write the ruleDoes this advice point at something on the screen, or assume something the screen cannot show? 93 of 116 pieces of advice were assumed.
  5. 05Re-checkAdvice that guessed fell from 80% to 68% on the screens we tuned on, but only from 92% to 88% on fresh ones.
One of the 41 · a Claude chat, before the fixes
A Claude chat reply explaining two collapsed audit findings, with copy, thumbs up, thumbs down and retry buttons below it.
FailTop gap: Confidence Visualization · Missing
“Not every chat needs a confidence score. The same goes for explainable AI: the audit treated this as the finished screen, when I can highlight any part and ask Claude to follow up. None of these gaps fit a simple chat.”

Imran's note on this audit

The same screenshot, after the fixes
Before Fail

Top gap: Confidence Visualization · Missing

After Pass

Top gap: Progressive Disclosure: collapsed audit-finding cards are unresolvable from within the chat

Guesses now go to a separate Worth checking list

  • Does clicking the right-arrow chevron on the collapsed-card bar expand the card content inline, or does it navigate away?
  • Does the 'Adaptive' model selector label explain what 'Adaptive' means on hover or click?

Across all screens, advice that guessed fell from 80% to 68%, and on fresh screens the fix had never seen, from 92% to 88%.

Read the full study

Get the one-page checklist

The five steps and a blank review table, as a PDF.

Free, and signs you up to AI UX Daily. Prefer a spreadsheet? Download the review sheet, no email needed.

Want to check every release without re-reading by hand? The AI Evals for Designers course shows how to hand your rule to an AI reviewer.