Introduction¶

People building Image/Video Generation Agents often encounter this situation: a user throws a line like “Rainy night Tokyo street, a woman holding a transparent umbrella, want a Kling 4.0 video prompt,” and the model treats it as a regular request, spitting out a prompt without negative prompts, parameter suggestions, or camera language. Conversely, developers don’t want to manually maintain a set of prompt specifications for every platform.

dsh-prompt-workshop (Prompt Workshop) solves this gap: select a target generator, input a rough idea, and the plugin provides the main prompt, negative prompts, parameter suggestions, multiple variants, and gap analysis. Below is an introduction to the plugin’s positioning, capabilities, installation, and usage.

What is it¶

One-sentence positioning: Native DeepSeek Harness (DSH) plugin that converts rough ideas into precise prompts usable by target generation models, complete with prompt quality scoring and image-to-prompt generation.

• Maintainer: 18518182595-cell
• Package Name / Version: dsh-prompt-workshop / 0.2.0
• License: MIT
• Category: Model Inference
• Source: GitHub README & package.json; Community Directory (skillhub.cn, independent site, no official affiliation with DeepSeek / Huafan).

It adopts DSH’s dual-half structure: the Host side provides tools and HTTP API, while the Browser side provides the input workspace and settings page. The built-in knowledge base is distributed with the package and only connects to the network during periodic collection.

Core Capabilities¶

Input Workspace¶

There is an entry button on the left side of the session input box; clicking it opens a floating panel. The flow is: Select target generator → Input idea → Generate → Copy / Insert into input box / Send directly. The result area also has one-click buttons for “More Concise / More Cinematic / More Photorealistic / Add Details” for iteration, so you don’t have to copy the entire prompt again.

19 Target Model Specifications¶

The built-in knowledge base covers syntax rules, parameter suggestions, and positive/negative prompt specifications for 19 target models, categorized into three types by the README summary table:

The README specifically points out the specifications for the two domestic models Doubao Seedream (image) and Doubao Seedance (video). The knowledge base research sources are concentrated in research/prompt-knowledge.md, and the README states there are 36 sources, including official docs and community cross-verification. Parameter suggestions are not strictly validated; entries marked “To be verified” follow the platform’s official documentation.

Multi-model Enhancement & Agent Tools¶

The enhanced model dropdown list directly lists all models configured in DSH “Settings → Models”: the full DeepSeek account line-up + any arbitrary OpenAI-compatible endpoints you added. The plugin does not save API Keys separately; it uses DSH’s existing model configuration. You can also set a default model in the settings page or select “Follow Current Session Model”.

After connecting to the conversation, there are three tools: prompt_enhance, prompt_score, and prompt_targets (including Doubao target model specification queries). In the conversation, simply saying “Turn this idea into a precise prompt” or “Why is this Midjourney prompt not effective?” will trigger the Agent to automatically call the corresponding tool; you don’t need to manually open the panel.

Quality Scoring¶

prompt_score uses a 10-dimensional rubric: Clarity of Intent, Specificity of Subject, Composition & Camera, Lighting & Color, Style Consistency, Motion & Temporal, Audio & Dialogue, Parameter Fit, Language & Grammar, Compliance & Copyright. This is particularly useful for video models—it covers temporal and audio dimensions, which is rare in general prompt tools.

Image-to-prompt¶

Upload or paste a reference image in the panel, and the plugin back-propagates the target model prompt that can “reproduce this image.” This step requires calling a large model that supports image input, otherwise, it will return a structured error.

Periodic Knowledge Base Update¶

The plugin checks for new news about relevant models/prompts online every N days (default 5 days) and distills it into an increment applied to the local knowledge base. Search prioritizes DeepSeek Search; if session credentials are missing, it automatically falls back to Bing RSS + official documentation seed pages, requiring no extra API key. Updates are divided into “Official Documentation Completion” and “New Changes,” generating a Chinese report each round. You can change the period in settings, click “Collect Now,” or click “Restore Built-in Knowledge Base.”

Saying “Update the prompt knowledge base” in the conversation also triggers this in the background. This path is mentioned in the README’s “Verified” section regarding test coverage: paths for Image-to-prompt and Knowledge Base collection without web/without attachment/without model failure all return structured errors.

Local Data¶

Settings and history are saved in ~/.dsh/prompt-workshop.json, and knowledge base increments are saved in ~/.dsh/prompt-workshop-knowledge.json. Upgrading the plugin won’t lose them. The built-in knowledge base is distributed with the package and only connects to the network during periodic collection.

Installation & Enablement¶

The README provides three ways; prioritize the first two:

Plugin Market (Recommended):

DSH → 插件 → 市场，搜索「提示词工坊 / Prompt Studio」

Or install from the Discover index.

GitHub:

dsh plugin --profile <profile> add github:<owner>/dsh-prompt-workshop

Local Link:

git clone <repo>
dsh plugin --profile <profile> add link:<absolute path to repo>

Desktop users can use the plugin manager link in “Settings → Plugins”. The README provides corresponding scripts/install-desktop.mjs and scripts/uninstall-desktop.mjs.

After installation, configure models and knowledge base in “Settings → Prompt Workshop” (or “Plugins → dsh-prompt-workshop”). Runtime requirements: node >= 22.13.0, dsh >= 0.2.1-rc.1, peer dependency @deepseek-ai/dsh-tools.

Typical Usage¶

The panel flow given in the README is 6 steps:

1. 打开任意会话，输入框左侧点入口按钮
2. 选「目标生成器」，按需选任务类型 / 风格 / 增强模型 / 变体数量
3. 在「粗糙想法」里写一句话，点「生成精准提示词」
4. 结果区复制 / 插入输入框 / 直接发送，或点快捷按钮继续迭代
5. 图片反推：面板里上传图片或直接粘贴进面板
6. 设置页设默认值、关系统提示注入、清空历史、调知识库更新策略

You can also state requirements directly in the conversation without using the panel. The README gives two examples:

把「雨夜东京街头，一个撑透明伞的女人」变成可灵 4.0 的视频提示词
帮我看看这段 Midjourney 提示词为什么效果不准

Scenarios & Notes¶

Suitable for these types of people:

• People calling multiple domestic image/video models (Doubao Seedream / Seedance, Jimeng, Kling, Hailuo) in DSH who don’t want to manually copy a set of syntax rules for every platform.
• People building Agent workflows who want prompt enhancement and scoring triggered in conversations rather than building strings themselves.
• People who need to batch produce multiple prompt variants of the same idea and then filter them.

Points to note:

• The plugin runs with the permissions of the current dsh process. Check the source code and license before installing (this repository is MIT, but it is still recommended to review it yourself).
• The API only accepts local loopback requests; access via LAN or reverse proxy is not available.
• Image-to-prompt and standalone API Key channels in the plugin were not present in v1; they were added in v0.2.0.
• The deep/light theme of the panel UI and floating layer positioning have not been pixel-perfectly confirmed; if the entry button next to the input box is missing after refreshing, refresh the page again (Ctrl+R).
• Parameters marked “To be verified” in the knowledge base are results of community cross-verification; the plugin only gives suggestions, and the platform’s official documentation is the final authority.
• The built-in knowledge base is distributed with the package, but periodic collection will search for external content online. Companies that mind can turn off auto-update first and switch to manual “Collect Now”.

Conclusion¶

The value of this plugin is not in “generating a beautiful prompt,” but in consolidating model specifications, scoring dimensions, and knowledge base updates in one place, and allowing them to be called by workflows via Agent tools and HTTP API. The README also lists candidates for Iteration 2 (multi-model comparison mode, prompt A/B testing, whitelisting for quality collection and change confirmation), so stay tuned.

Directory Page: https://www.skillhub.cn/plugins/18518182595-cell/dsh-prompt-workshop

GitHub: https://github.com/18518182595-cell/dsh-prompt-workshop