Foreword¶

A common issue in front-end development within DeepSeek Harness (DSH) is that code runs and passes static checks, but upon deployment, users find the workflow cumbersome, information hierarchy unclear, and key actions buried deep. Tools like axe and Lighthouse can catch “absolute rules” such as contrast ratios or missing alt text, but they struggle to determine whether something is problematic “for a specific user type.” For example, a confirmation dialog before deletion might protect occasional users but burden operators who process hundreds of records daily. Without context about the target users, many “UX issues” are essentially undefined.

Here, we introduce the community plugin dsh-user-experience (maintained by DietCokewithSugar, 19 stars on GitHub, categorized under model inference). It uses target user personas as a baseline to scan front-end source code—such as React or Vue 3—during development, pinpointing experience issues and providing reviewable optimization suggestions, rather than waiting for user feedback post-deployment.

What This Is¶

dsh-user-experience is a UX review plugin for DeepSeek Harness. In short, it lets AI simulate target users to walk through the product, discovering experience issues during development and offering specific, locatable, and reviewable optimization suggestions.

Unlike common CLI-style tools, its README explicitly describes it as a pipeline rather than a standalone command-line tool. There’s no fixed command like /ux; you initiate a review using natural language or it triggers automatically after changes to front-end files.

If the project hasn’t defined user personas, the plugin infers 1–3 draft personas based on the README and routes. After your confirmation, it proceeds with the review—no extra configuration commands needed.

Core Features¶

Persona-Driven Review Logic¶

Each finding is tied to an active target user persona. The review report starts with natural language: which page, what happened, and the impact; technical details like file paths and rule IDs are collapsed in a “Technical Details” block and can be copied with one click for further processing by other AI. Decisions also support natural language, such as “The second one isn’t an issue” or “Ignore anything below level three.”

Supported Inputs and Evidence Levels¶

Each finding is annotated with an evidence level: static, rendered, or interactive. Without browser capabilities, the review continues based on static evidence without pretending to “see the page.” Conclusions about layout density, visual language, and primary actions require rendered evidence; redundancy findings require interactive persona reviews.

It explicitly does not support—and will not make low-quality guesses about—stacks like Svelte, Vue 2, or mini-programs (.wxml).

Natural Language Entry and Auto-Review¶

You can initiate a review using natural language, e.g., “Check the checkout flow” or “Is this page usable?” If personas are missing, it first drafts them and asks for confirmation, then immediately starts the review.

After editing front-end code, the review runs automatically: it no longer asks about personas repeatedly but only interrupts you for level-one or level-two issues. After confirming a finding, the plugin generates a task prompt for coding AI, describing the observed phenomenon rather than prescribing code changes, and allows text modifications.

Differences from Generic Accessibility Tools¶

axe and Lighthouse verify objective, rule-based criteria. This plugin takes “for whom it’s a problem” as a premise, focusing on relative experience judgments during development. The two are complementary, not replacements.

Installation and Enablement¶

Run the following commands in DeepSeek Harness (SkillHub directory page matches the GitHub README, current version 0.4.2):

dsh plugin --profile web add dsh-user-experience@0.4.2

It’s recommended to pin the version number and avoid using @latest. The README notes that pnpm 11 may delay releases within 24 hours, so @latest might not resolve the package in a new profile. For updates, check the npm version list.

After installation, refresh the page—usually no restart is needed. Only if the marketplace indicates hot-reloading is unavailable should you restart or reload the web profile.

Migration from github: Installation Method¶

dsh plugin add github:DietCokewithSugar/dsh-user-experience is no longer available: the repository no longer commits lib/ build artifacts, and pnpm by default blocks build steps. Use the package name without the github: prefix:

dsh plugin --profile web add dsh-user-experience@0.4.2

If the marketplace still resolves a Git reference, remove the old github: dependency line from the profile’s package.json, delete node_modules and pnpm-lock.yaml under that profile, and re-run the installation command.

Typical Usage¶

1. Initiate Review with Natural Language¶

In the Harness chat, state your intent directly, such as checking if a page or flow is usable. If no personas exist, the plugin displays 1–3 inferred target users; after confirmation, it reviews from those perspectives.

2. Review the Report and Make Decisions¶

The report card first provides the page, phenomenon, and impact description. Expand technical details to view static, rendered, or interactive evidence when needed. If you disagree with a finding, click a button or explain in natural language.

3. Hand Off to Coding AI After Confirmation¶

For confirmed issues, copy the plugin-generated task prompt to another session or tool, allowing AI to propose modifications with full project context. The plugin itself does not automatically change code.

4. Auto-Review After Code Changes¶

After saving front-end files, the review runs in the background, only alerting for high-priority (level-one, level-two) issues without disrupting the regular coding rhythm.

Use Cases and Notes¶

Who It’s For: Teams developing React (TypeScript/JavaScript) or Vue 3 front-ends under the DSH web profile, who want to discover experience risks in workflows, layouts, and primary actions from a persona perspective before merging or releasing. When the current Harness session can launch the app, it can also combine browser evidence for rendering and interaction layer verification.

Permissions and Security: The plugin runs with current DSH process permissions, accessing source code and session context within the workspace. Before installation, read the repository source code and LICENSE. The README also includes security notes in the installation section (#installation).

Ecosystem Note: DSH follows an “everything is a plugin” philosophy; SkillHub (skillhub.cn) is a community directory for Chinese users, not officially affiliated with DeepSeek or High-Flyer. This article’s installation commands are based on the directory page and GitHub repository.

Capability Boundaries: CSS analysis can provide inspection clues, but conclusions about whitespace, hierarchy, and visual quality won’t be asserted without real rendered routes. Unsupported tech stacks are honestly labeled without low-quality guesses.

Conclusion¶

dsh-user-experience threads target user personas, source code static analysis, and optional browser evidence into a development-stage UX review pipeline: triggered by natural language or code changes, reports are readable and decidable, and confirmed findings can export task prompts for further iteration. If you’re doing front-end work in Harness and care about “real-user usability,” try installing it with a pinned version.

• SkillHub Directory Page: https://www.skillhub.cn/plugins/DietCokewithSugar/dsh-user-experience
• GitHub Repository: https://github.com/DietCokewithSugar/dsh-user-experience