Introduction

When developing agents with DeepSeek Harness (DSH), adjusting frontend UI usually requires developers to manually refresh the page and compare results. Manual intervention is not only inefficient but also prone to overlooking details. screenshot-review solves this problem by automating the screenshot, evaluation, and correction workflow.

What Is This?

This is a DSH skill: it lets the model take screenshots, view the images itself, and modify the code, iterating repeatedly until the frontend result meets the requirements.

  • Maintainer: wenyixiaoqingnian
  • Core value: runs the screenshot review loop (modify code → refresh page → take screenshot → view screenshot → modify again) until the frontend result meets the requirements.

Core Features

  1. Screenshot review loop: the model performs code modifications, then refreshes the page and takes a screenshot. The model uses read_image to view the screenshot; if it is unsatisfied, it continues modifying the code until the visual appearance of the page meets the requirements or the maximum number of iterations is reached.
  2. Depends on dsh-browser: The skill internally calls the browser_open, browser_navigate, and browser_screenshot tools provided by the dsh-browser plugin.
  3. Self-correction: The model has the ability to view screenshots and correct its code autonomously.

Installation and Enablement

Before installation, make sure the DSH Harness environment is installed.

  1. Install the browser screenshot plugin (provides browser_* tools)
dsh plugin --profile <你的profile名> add dsh-browser
  1. Install this skill package (includes dsh.bundle and automatically registers the skill)
dsh plugin --profile <你的profile名> add github:wenyixiaoqingnian/screenshot-review

After installation, restart dsh web (or reopen the conversation) to activate it.

Typical Usage

In the conversation, send the following instruction directly to the model:

Open http://localhost:3000, check the result, and help me make it look better. After making changes, take a screenshot to confirm, and keep iterating until you are satisfied.

When the model detects keywords such as “look at the result / screenshot review / iterate until satisfied”, it will automatically load this skill and execute the screenshot iteration loop.

Application Scenarios and Cautions

  • Dependencies:
    • The dsh-browser plugin must be installed.
    • The model must support image input (such as deepseek-v4-flash or GLM-4V) in order to call read_image and view screenshots.
  • Iteration limit: The system is configured with a maximum of 5 iteration rounds. The workflow stops when the model reaches the limit or confirms it is satisfied.
  • Permissions and security: The plugin runs with the permissions of the current DSH process. Before use, it is recommended to review the source code and license (MIT).

Summary

screenshot-review delegates the entire “modify code — view image — modify again” frontend UI tuning process to the model for autonomous execution. It relies on the browser operations provided by dsh-browser, combined with a model capable of visual understanding, to achieve self-consistent iteration of the frontend result.

  • GitHub: https://github.com/wenyixiaoqingnian/screenshot-review
  • Catalog: https://www.skillhub.cn/plugins/wenyixiaoqingnian/screenshot-review