Objective
Your mission
List the AI's claims about a build, test each one, test edge cases, and fix any failures. Output: verification_report.md.
You need: Any AI-built project you can run and interact with.
Time: 25 minutes.
Step 1: list the AI's claims (3 min)
After every build, the AI says things like "I've added the feature, the form validates input, the app is mobile-responsive." These are claims. List them.
Write in your file:
## AI's Claims
1. [claim 1 — e.g., "form validates input"]
2. [claim 2 — e.g., "calculation updates live"]
3. [claim 3 — e.g., "app is mobile-responsive"]
...
Step 2: test each claim (10 min)
For each claim, test it manually:
"Form validates input" → Enter an empty form. Enter invalid email. Enter valid email. Does it behave correctly for each?
"Calculation updates live" → Type a bill amount. Change tip percentage. Change people count. Does it update without clicking anything?
"Mobile-responsive" → Open on your phone or use browser dev tools mobile view. Does the layout adapt?
Write in your file:
## Claim Testing
1. [claim] — [pass/fail] — [what happened]
2. [claim] — [pass/fail] — [what happened]
3. [claim] — [pass/fail] — [what happened]
Step 3: test 5 edge cases (7 min)
The AI tests the happy path. You need to test edge cases:
- Empty input — what happens?
- Zero (0) — does it handle $0?
- Negative number — does it accept negative values?
- Very large number — does it handle $1,000,000?
- Special characters — what happens with
<script>alert('test')</script> in a text field?
Write in your file:
## Edge Case Testing
1. Empty input: [result]
2. Zero: [result]
3. Negative number: [result]
4. Very large number: [result]
5. Special characters: [result]
Step 4: fix failures (5 min)
For any claim that failed or edge case that broke:
Write in your file:
## Fixes Applied
1. Issue: [description]
Correction to AI: [what you asked]
Result: [fixed / still broken]
2. [repeat for each issue]
The trust spectrum
You don't need to distrust the AI — you need to verify it. Distrust means assuming everything is wrong. Verification means checking everything regardless of whether it seems right. The AI is trustworthy enough to build with, but not trustworthy enough to ship without testing.
The AI says "the form validates input." What does verifying that claim mean?
Claims are tested, not trusted: empty fields, invalid email, zero, negatives, huge numbers and a script tag. The AI tests the happy path; edge cases are yours.