Generate a precise SVG layout of text/numbers first, then pass it as an image input to Gemini or ChatGPT to paint a styled visual on top.
Key Takeaways
The underdrawing method splits work by tool strength: SVG/HTML handles math and layout deterministically; image models handle visual style.
Both Gemini 3.0 Pro and ChatGPT Images 2 fail a 50-stone spiral board prompt without this technique; both succeed with it.
Workflow is two steps: generate an SVG wireframe via Claude Code, then send SVG image plus style prompt to an image model with image+text input support.
The method outperforms native text rendering in current state-of-the-art models and is expected to be absorbed into model pipelines soon.
Results are not perfect every time; the author notes occasional failures even with the underdrawing approach.
Hacker News Comment Review
Minimal discussion so far; consensus is that the trick is obvious in hindsight but genuinely useful and not yet widely known.
Notable Comments
@sparuchuri: “duh, why didn’t I think of that” – calls it a clear practical fallback when imagegen text rendering fails.
