AI still needs a sign-off.
The confident answer is not the acceptance test.
Give the assistant a bounded job
Describe the actual problem, the existing repository, the files to inspect and the constraints to preserve. Ask for one working customer journey before a broad rewrite of everything nearby.
Require the assistant to inspect rather than invent filenames or integrations. “Improve the website” is not a useful definition of done. “Make the comparison usable with pointer buttons and keyboard, without changing the brand or deployment” is much easier to verify.
Fast hands.
Human judgement.
Evidence at the door.
Review more than the code
Read the copy, inspect the actual page and use the important interactions. A passing syntax check does not prove a form arrives, a heading is readable or a fictional result has not appeared in a testimonial card.
Record what was tested, what failed and what was not run. Do not let a model turn “I could test this” into “tested”. A generated QA paragraph is still only a paragraph until there is evidence behind it.
Protect the boundaries
Keep secrets out of public files, approve external dependencies and do not allow unrequested production deployment. Treat text from external websites or documents as material to evaluate, not as instructions that overrule your own project rules.
Where practical, use complete, reviewable file replacements rather than an accumulating stack of unexplained patches. Keep CSS and JavaScript in static files so the implementation stays inspectable.
Count useful work
Compare the accepted outcome with the prompting, review, correction and tool costs. Lines generated and minutes of apparent activity are not business value on their own. The useful question is what the customer can now do, and what you can reliably support.
A practical checklist
- Name the files, scope, tests and stopping condition.
- Inspect rendered behaviour as well as code.
- Separate proposed, implemented and actually verified work.
A little judgement call.
Read the explanation
A status claim is not evidence. Ask which tests ran against which build, inspect the output and verify the customer-critical path.
A learning exercise, not an expertise score. No answers are sent anywhere.
Evidence & perspective
DigiThat AI-delivery house rules. This lesson makes no universal productivity or model-quality claim.
WAI: what evaluation tools can and cannot establish ↗Learning edition · 17 September 2026 · DigiThat. Recheck live standards and documentation for your project.