How do you accurately implement a messy Figma design with AI coding tools?

I use Codex to implement frontend pages from Figma designs. When the Figma file is well structured, the workflow is straightforward. But some real project files have poorly named layers, inconsistent grouping, and unclear spacing or alignment rules.

My current workflow is:

• Give Codex as much Figma context as possible.
• Use detailed prompts that discourage unnecessary absolute positioning and favor normal document flow, Flexbox, and Grid.
• Provide screenshots as an additional visual reference.
• Review screenshots of the implementation against the original design and iterate on the differences.

The weak point is still the messy Figma source: AI can interpret component boundaries, decorative layers, and layout rules differently from what the designer intended.

How do you make this workflow more reliable?

• Do you prioritize screenshots, Figma metadata, or manual design inspection?
• How do you distinguish real UI from device frames or decorative content?
• How do you decide component boundaries when the Figma hierarchy is not trustworthy?
• What tools or visual-comparison workflow help you move toward pixel-level accuracy?

I would love to hear practical workflows from frontend developers using Codex or similar AI coding tools.