Coding Agent Acceptance Criteria Checklist

A coding agent acceptance criteria checklist is the minimum contract between the human and the model before code changes are accepted.

The first item is the behavioral target. “Fix the bug” is too broad. “When the user opens the node page on mobile, the body text should wrap without horizontal scrolling” is usable because it names the surface, the user-visible behavior, and the expected result. The second item is scope. The model should know which files, routes, or components are likely involved and which areas should not be touched. Scope protects the codebase from broad refactors disguised as fixes.

The third item is verification. For backend work, that may be a test command, API response, migration dry run, or log check. For frontend work, it may be a browser screenshot, responsive viewport pass, accessibility check, or canvas pixel sanity check. For content or SEO work, it may be a canonical URL, title, meta field, sitemap entry, and public GET response. Verification must be written before the model declares success.

The fourth item is failure handling. If a test cannot run, the agent should say why, what it tried, and what risk remains. A failure should not be hidden behind confident prose. The fifth item is patch discipline: changed files, reason for each change, and any unrelated dirty state that was left alone.

This checklist makes Claude vs GPT/Codex comparisons more concrete. A model that writes a better first answer but skips verification may be less useful than a model that makes a smaller patch and proves it. The checklist does not remove judgment; it gives judgment a record.

Coding Agent Acceptance Criteria Checklist

// COMMENTS

ON THIS PAGE