DEV Community

Discussion on: I changed how I write acceptance criteria, and my AI agent stopped building the wrong thing

Collapse
 
kunalsharda profile image
Kunal Sharda

Exactly that. "Fast" and "handles errors" are the two phrases that have wasted the most agent-tokens in my codebase.

The riff I'd add: the discipline gets harder once you leave backend territory. "p95 < 200ms" — clean. "Returns 4xx with json body" — clean. But "the form feels responsive" or "the empty state explains what to do" — those resist numbers, and you end up writing them as either a precise string ("show empty-state copy: 'No items yet. Add your first one.'") or you punt them to a Figma reference.

The hidden second unlock for me: once every criterion is checkable, the agent can verify its own work before opening the PR. The same definition you use as input becomes the test it runs as output. Closes a feedback loop that used to need a human in the middle.

What do you do for the criteria that resist clean numbers?

@uzoma_uche_3ec83974b4a8a5