Reliable node-tree traversal in Figma via Claude Code — anyone solved this?
I'm a design director doing heavy Figma automation work — copy migrations, design token audits, making sure variable bindings are actually correct across hundreds of components. The kind of work where "walk every node and check every property" is the whole job.
I've tried Figma's official MCP, Figma Console MCP, and a Figma CLI wrapper. All have the same failure mode: disconnections, slow round-trips, and — worst — silent incompleteness. Claude will report "done, verified" after a full audit and then I find a dozen nodes it never touched. On a recent wireframe-to-component migration it skipped entire frames and still claimed full coverage.
I ended up building my own CDP-based CLI (drives Figma Desktop directly, no plugin/API key) with token import/export, component scaffolding, and screenshot-based self-verification specifically because I needed the AI to prove its work rather than assert it. It's better, but I'm still fighting the same core problem: getting reliable, complete traversal of a large node tree without it losing track partway through.
Anyone found patterns that actually hold up at scale? Specifically:
- Forcing verification against actual node counts instead of taking Claude's "done" at face value
- Chunking/batching strategies for large trees that don't lose state between batches
- Anything that catches silent skips before they ship
Want to spend less time babysitting and more time doing the actual design work.