No More Screenshots and Red Boxes When Asking an Agent to Revise a Design

Baoyu's baoyu-design has been upgraded to v2.1. This time the added feature is comments.

On a design page or slide, directly mark an element and write a comment for it. The Agent can read the comment and reply.

It solves the most awkward step in getting AI to revise a design: not being able to say clearly what to change.

Anyone who has worked on design files has probably run into this. You ask the Agent to change a button's color, and it changes the wrong place because there are three or four buttons on the page that look almost the same. Take a screenshot, draw a red box, paste it back, and explain, "It's the one on the left." After a few rounds, you've spent more time describing than designing.

Now instead, you click the element in the preview and write, "Change this button to the primary color." The comment goes into comments.json, the Agent reads it, makes the change, and replies to you.

A user named Panda went through the repo and explained the mechanism clearly: comments are stored in the project's local comments.json, and the Agent reads and replies through artifact.mjs. When it runs comments --wait, an extra Send to agent button appears in the editor; once you've marked something, send it straight over. Resolved comments are not listed by default; only with --all are they brought out.

This detail is small, but practical. Handled comments won't pile up and get in the way of the next round, but you can still dig up old history when needed.

More noteworthy is where comments are stored. They sit together with the project data and become part of it. Open the project three months later and you can see why the spacing was changed from 16 to 24. The design rationale no longer gets scattered across chat logs or someone's memory; it travels with the files.

One user said annotations used to deserve only the name "complaints." It's a bit self-deprecating, but accurate. Old annotations were for people to read and then discard. Now annotations are both input for the Agent and project archives.

A quick note on what v2 itself does. It fully ports over Claude Design, Claude Slides, and Claude Dashboard, with an experience nearly identical to the original—except for comments. It can run on any Agent and any model, no longer dependent on Claude Artifacts. Fully open source, MIT license.

How Claude Dashboard looks when run locally:

Claude Slides running in Codex can be edited with the built-in editor and exported as editable PPTX:

Import a design system with the Claude Design System skill to keep the style consistent:

Running Claude Design locally keeps files and the project together; no need to import/export between online Artifacts and a local Project. Edit directly in the editor and sync changes back to local files, or initiate edits from the Agent side—two-way sync:

One take: the comments feature, which looks the most unassuming in this update, may be the most useful one.

Generation quality is the model's job, and it will keep improving. The question of "how to accurately convey human intent to the Agent" won't automatically disappear just because models get smarter; the bottleneck is the human cost of expression. The comments feature works on exactly that side.

Users' follow-up questions also show they're already using it as a collaboration tool. Some ask whether a comment is automatically marked resolved after the Agent makes the change or has to be closed manually; others ask whether you can select multiple spots and share one comment, for example when several headings all need a color change. The materials don't give a definitive answer; you'd need to check the repo or actually run it.

Another user put it bluntly: clients don't know how to write prompts, but they can point at the page and say, "This is wrong."

This may be the real user profile for this feature. People who can write prompts don't need it that much to begin with. People who don't write prompts can express themselves by pointing at the page. Where the barrier to expression is lowered, people are willing to use it.

Repository here: https://github.com/jimliu/baoyu-design

Install:

npx skills add JimLiu/baoyu-design