Editable Visual Design skill makes coding agents output posters as editable HTML with layers and real text

Junyan Ye and eleven coauthors, most of them at Sun Yat-Sen University, released Editable Visual Design on 4 September 2026. The GitHub repo yejy53/Editable-Design ships two Codex Skills and links to an arXiv paper. The lead skill has a coding agent produce posters, infographics and marketing campaigns as an HTML file where each element sits in its own layer, and the text stays real text.

What the skill actually produces

The output is an HTML file where each element sits in its own layer. Text is real text a user can select and rewrite in the visual editor the skill ships alongside. Images are separate assets the agent calls a diffusion model to generate on demand, then places into the HTML. The repo also includes a sibling skill, html-to-pptx, that converts the same HTML into an editable PowerPoint file when the source structure supports it.

The authors say the coding agent runs an "imagine first, then act" loop: a vision-language model plans the layout and aesthetic, the agent writes native HTML and CSS, calls the image model for isolated assets, renders the design, looks at the render, and refines. A feature they call Agent Design Replay records the reasoning trajectory so a user can step through the choices the agent made.

What the authors compare it to

The paper (arXiv 2609.04034, revised 4 September 2026) opens by naming GPT-Image-2 and Nano-Banana as the current visual-expressiveness leaders and says both "inherently yield flattened bitmaps with error-prone text, precluding layer-wise post-editing." The authors position their approach against that: a local change no longer means regenerating the whole image. The paper reports qualitative validations on posters, infographics and academic posters, and does not publish a benchmark table.

The README recommends running the skill on GPT-5.6 with high reasoning effort or above. The gallery in the repo shows fourteen prompts across campaigns, information design, text-led design, posters, academic posters and art design, each with the final render, an editable version and a replay video.

The traction and the caveat

The yejy53/Editable-Design repository was created on 24 August 2026 and has 624 stars and 55 forks as of the time of writing, per the GitHub API, over its first 15 days. The lead author Junyan Ye is a researcher at Sun Yat-Sen University. The code is released under Apache 2.0, and the repo takes commits in the last few hours. There is no independent evaluation yet, so the "editable at production quality" claim is the authors' own. Anyone considering this for real client work should judge the gallery against a familiar design task before adopting it.

For a team that already runs a coding agent to produce marketing and product visuals, the difference is that the copy on a poster, the price on a product card, or the label on a chart can be changed without asking the agent to regenerate anything. That is the workflow change worth testing.

Source

• Repository: github.com/yejy53/Editable-Design
• Paper: Editable Visual Design, arXiv 2609.04034