Repeated false completion claims during Figma-to-code implementation

I asked Codex to implement a UI component exactly according to a provided Figma design. Despite having access to the Figma node and a clear requirement for pixel-perfect implementation, the agent repeatedly claimed that the result was correct before performing an adequate visual comparison.

Problems observed

• Missing elements such as the description and action button were initially overlooked.
• RTL alignment was implemented incorrectly.
• Card text, spacing, underline width, image positioning, and CMS titles did not match the design.
• The agent used an incorrect check-circle icon instead of the clearly visible left-chevron icon.
• When I explicitly asked whether the icon was correct, the agent incorrectly confirmed that it matched Figma.
• I had to compare the implementation with the Figma design repeatedly and point out obvious visual differences one by one.
• The agent declared the work complete several times while visible mismatches still remained.

This caused significant time loss and reduced trust in the agent's verification process.

Expected behavior

When a Figma design is provided, Codex should compare the rendered page directly against the design at the same viewport. It must verify text, icons, RTL alignment, spacing, dimensions, borders, radii, image cropping, and responsive behavior before declaring completion.

If Figma metadata or an exported asset conflicts with the visible design, the rendered Figma screenshot should be treated as the final visual reference. Codex should clearly state that the implementation is incomplete until all visible differences have been resolved.

Start with the Figma-backed UI implementation and the completion-verification flow, using the rendered page and Figma design at the same viewport. Done means visible text, icons, RTL alignment, spacing, dimensions, borders, radii, image cropping, and responsive behavior have been compared and the agent does not claim completion while mismatches remain.