Claude Design Runs in DeepSeek Harness, No More Screenshots and Red Boxes for Revisions
Baoyu made a DeepSeek Harness plugin that lets Claude Design annotate and comment directly on design pages and slides. The Agent reads the comments and then replies and makes changes. Comments are stored together with project data, reducing the problem of losing context when switching windows.
Claude Design can now run inside DeepSeek Harness. The person behind this is Baoyu. It is a dsh plugin, not an official integration between the two companies.
Plugin address: https://github.com/JimLiu/baoyu-design/tree/main/dsh-plugins
Original tweet: https://x.com/dotey/status/2116336030584574076
The main video demonstrates three actions: annotating the canvas, sending it to the Agent, and the Agent replying in a thread.
If you're not familiar with this toolset, here's some context: baoyu-design v2 ported out Claude Design, Claude Slides, and Claude Dashboard. It is open source and can run on any Agent and any model, no longer relying on Claude Artifacts. v2.1 added commenting. Now the dsh plugin connects this path to DeepSeek Harness and preserves the closed loop of annotation, comments, and revisions.
The most annoying part of changing a UI is often describing location. The button in the upper left, below the title, that blue card—after several rounds of back-and-forth, the Agent still may not find the right spot.
The comment feature pins feedback to elements. You can mark elements on design pages and slides and add comments; the Agent reads the comments and can also reply in the comments. All comment data is stored together with project data, becoming part of the project's data assets.
A user looked through the README and said comments are stored locally in the project's comments.json, and the Agent reads and replies through artifact.mjs. When running comments --wait, an extra Send to agent button appears in the editor; resolved comments are not listed by default, and only with --all are they included.
What's more worth looking at here is whether comments travel with the project; the model name comes later. A user said switching models is not the point—whether the annotations are still in the project is. If the points marked in the previous round don't carry over, the Agent in the next round is essentially guessing again.
When multiple people edit the same element at the same time, ordering may become a problem before the model name does. User zrqgood also asked a specific question: are comment anchors bound by element coordinates or by the page DOM tree? Will historical comments be misaligned after the page structure is rearranged? Is there a conversion layer for pixel offset under 4K and 1080p rendering? The source material does not provide answers, but these questions directly affect team use.
For people building B2B products, the value is not in saving money. When a customer's intranet only approves domestic models, the design draft step doesn't have to go outside. Annotations and comments can be written back; a customer pointing at the page and saying 'change the color here' is more reliable than relaying it to the Agent. User Light@白酒云 mentioned this scenario.
Usage is not complicated. The plugin is in the dsh-plugins directory of the baoyu-design repo: https://github.com/JimLiu/baoyu-design/tree/main/dsh-plugins. The README adds documentation for DeepSeek Harness Plugins. The author says writing a dsh plugin is simpler than expected: give the official docs to Opus 5.5, state the requirements, and it can be done. Official docs: https://deepseek-harness.github.io/deepseek-harness/en/develop/basic/. He also tried DeepSeek 4.1 Flash, and the results were good.
DeepSeek 4.1 Flash generated results:
If you plan to put it into a team workflow, test three things first: whether comments travel with the project, whether comments become misaligned when the page structure changes, and how conflicts are handled when multiple people annotate at the same time. Whether to switch models can actually be decided later.
发布时间: 2026-11-01 03:35