The design agent that ships editable source, not just a PNG
Generative image models hand you a flat PNG with mangled text; this research project has a coding agent build posters, infographics, and paper diagrams as semantic HTML with real, editable text and independent layers.

What it does
Editable-Design is a research project (paper: arXiv 2609.04034) delivered as a set of Codex Skills. You describe a poster, infographic, marketing campaign, or paper workflow diagram; one persistent agent plans the visual composition, builds it as semantic HTML with real text and independent image assets, renders it, and repairs what’s broken. Alongside the PNG you get the editable source: a visual editor, a layer breakdown, and an Agent Design Replay showing how the artifact was assembled. A companion skill converts compatible HTML into editable PowerPoint.
The interesting bit
The core bet is that visual design is a coding problem, not a diffusion problem. Because the artifact is code, every headline and price stays a string you can edit, and the agent can check its own homework — rendering checks and format-specific contracts — then repair failures instead of re-rolling a fresh image. The Replay is the unusual part: the creation path is treated as a first-class output, not discarded scaffolding.
Key highlights
- Three skills:
editable-design(posters, infographics, campaigns → HTML, PNG, editor, layer breakdown, replay),paper-fig(method descriptions → editable PPTX diagrams), andhtml-to-pptx(existing compatible HTML → editable PPTX). - Real text and independent layers — change a price or a headline without regenerating the whole design.
- One agent plans, builds, renders, and repairs; editing and verification happen locally, with PPTX conversion running through Playwright Chromium in an isolated Python environment.
- Gallery of thirteen examples spanning campaigns, information design, text-led design, posters, and art design.
- Backed by a paper on arXiv; the README notes Paper Fig was developed with GPT-6.
Caveats
- As documented, it’s built for Codex — the skills install under
~/.codex/skills/, and the README offers no path for other agent setups. - The README recommends
highreasoning effort or above for complex diagrams and layouts, so this is not a cheap-run tool. - The authors are blunt that output still needs human review for scientific accuracy — don’t auto-trust the paper figures.
Verdict
Worth a look if you produce posters, infographics, or paper figures and are tired of AI output you can’t edit — provided you already run Codex. If you’re on another agent stack or want a standalone design app, the README doesn’t meet you there.
Frequently asked
- What is yejy53/Editable-Design?
- Generative image models hand you a flat PNG with mangled text; this research project has a coding agent build posters, infographics, and paper diagrams as semantic HTML with real, editable text and independent layers.
- Is Editable-Design open source?
- Yes — yejy53/Editable-Design is open source, released under the Apache-2.0 license.
- What language is Editable-Design written in?
- yejy53/Editable-Design is primarily written in Python.
- How popular is Editable-Design?
- yejy53/Editable-Design has 1.2k stars on GitHub.
- Where can I find Editable-Design?
- yejy53/Editable-Design is on GitHub at https://github.com/yejy53/Editable-Design.