// the find
JimLiu/baoyu-design
Run Claude Design locally as an Agent Skill — Cursor, Claude Code & more. Produce polished UI mockups, prototypes, decks & wireframes as self-contained HTML, without claude.ai/design. Best with Opus 4.8.
baoyu-design packages Anthropic's Claude Design methodology as a portable agent skill you can run in Cursor, Claude Code, Codex, or any file-capable agent. It produces self-contained HTML mockups, prototypes, decks, and wireframes without needing claude.ai/design. The target user is a developer or designer who already has a capable local agent and wants design-quality output inside their editor workflow.
The gen-pptx pipeline is genuinely clever: rather than trying to parse HTML, it renders in a real headless Chromium, walks the live DOM for computed styles and pixel-precise bounding boxes, and writes native PowerPoint shapes and animations — so animations actually round-trip to the Animation Pane rather than getting baked into screenshots. The offline Figma .fig decoder is a real differentiator; sparse-checkout of only needed paths before a full clone shows the author thought about the slow path. The starter-component library (device shells, deck-stage, chart-stage, animation primitives) saves the agent from reinventing the same scaffolding every run. The design-system binding via _d_meta.json is the right call — it keeps projects self-contained and reproducible without reaching outside the folder at render time.
The whole thing is a prompt engineering project, not a library — quality depends almost entirely on the underlying model, and the README is honest that Opus 4.8 is needed for best results, which is not cheap. The PPTX editable export path does a recursive DOM walk and then maps CSS to PowerPoint concepts by hand; CSS layouts that don't translate cleanly (grid, subgrid, complex flex nesting) will silently degrade and there's no escape hatch other than the screenshot mode. There's no sandboxing: the agent writes files to your repo and starts a local HTTP server — nothing stops a malicious prompt from directing the skill to write outside designs/. The upstream-sync mechanism (provenance.json, apply-report.json) exists but there's no automated process to pull changes from claude.ai/design when Anthropic updates the methodology, so the community fork will drift.