メインコンテンツへスキップ
Y

dsh-ai-workflow-kit

yyyjjjqqq714-spec/dsh-ai-workflow-kit

340-entry code-animation style library, 8 prompt templates and 5 renderable code templates in a sidebar panel, plus agent tools (awkit_search / awkit_get_style / awkit_render_template) so the model can pick a style and render a template during a turn; content ships inside the package, no extra download.

インストール

dsh plugin --profile web add github:yyyjjjqqq714-spec/dsh-ai-workflow-kit

README

AI Workflow Kit for DeepSeek Harness

中文说明

A sidebar panel that puts a 340-entry code-animation style library, prompt templates and renderable code templates one click away from the chat — and exposes the same library to the agent as tools, so it can search a style and write the implementation prompt for you.

Everything ships inside the package: no extra download, no network call, no absolute path baked into your machine.


Why this exists

Prompt libraries usually live in a folder you forget the path of. This one is a plugin: the library is browsable as thumbnails, searchable by id / name / tag, and callable by the agent during a normal conversation.

  • 340 video styles across 10 series (code samples, cartoon, motion graphics, Chinese ink, Oscar-film palettes, code-video, …), each with the original implementation prompt.
  • 8 prompt templates for storyboards, title cards, style mixing, style → code, tutorials and scripts.
  • 5 code templates (Fastify route, React component, Zod schema, FastAPI router, argparse CLI) with variables and one-click rendering.
  • Storyboard forge: pick a style, type a topic, get a complete director instruction with platform rules baked in.

Install

From the plugin market, or:

dsh plugin --profile web add dsh-ai-workflow-kit

Pin a release tarball instead:

dsh plugin --profile web add "https://github.com/yyyjjjqqq714-spec/dsh-ai-workflow-kit/releases/latest/download/dsh-ai-workflow-kit.tgz"

Requires DeepSeek Harness >= 0.1.7. Restart DSH after installing a replacement build; a first install can activate through HMR.

Behind a proxy — DSH's plugin manager shells out to pnpm, and pnpm does not read the OS proxy settings. If the install hangs on ETIMEDOUT while your browser can reach GitHub, pass the proxy explicitly:

$env:HTTPS_PROXY = "http://127.0.0.1:7897"   # your own proxy
dsh plugin --profile web add "https://github.com/yyyjjjqqq714-spec/dsh-ai-workflow-kit/releases/latest/download/dsh-ai-workflow-kit.tgz"

Use

Sidebar — click AI 工作流 to open the panel. Three tabs, one search box (/ focuses it).

  • 视频风格 — thumbnail wall grouped by series. Click a card for the full prompt, a "copy prompt" button, and 生成分镜指令 which turns the style plus a topic / platform / duration into a ready director instruction.
  • 提示词 — 8 templates, copy or send straight into the conversation.
  • 代码模板 — fill the variable form, watch the rendered source update live, copy it or send it to the chat.

Slash command

/awkit                 overview + help
/awkit 水墨            keyword search, shows the top hit's prompt
/awkit S1-A3           exact id
/awkit prompts         prompt inventory
/awkit templates       code template inventory

Agent tools — the model can call these during a normal turn:

ToolWhat it does
awkit_searchsearch styles / prompts / templates
awkit_get_stylefull style prompt by id
awkit_render_templaterender a code template with variables (dry-run by default)
awkit_overviewinventory of the whole library

Typical ask: "用 S1-A3 这个风格,帮我做一条 45 秒抖音讲别熬夜的分镜" — the agent finds the style, pulls the prompt, and produces the storyboard.

Privacy and permissions

  • Content is read from the package itself (content/), through node:fs read-only calls.
  • The only HTTP surface is a same-origin prefix route /awkit/api used by the panel in your own browser.
  • No telemetry, no outbound network requests, no shell.
  • Writing is opt-in and single-purpose: awkit_render_template writes only when the caller passes write: true and a targetDir, and refuses paths that escape that directory.

Structure

index.js            host entry — assembly only
lib/content.js      content layer: load, cache, search, error codes
lib/routes.js       same-origin REST API (old and new routes both supported)
lib/tools.js        agent tools
lib/commands.js     slash commands
client.js           browser half: panel, detail views, forge, renderer
content/            340 styles + thumbnails, 8 prompts, 5 templates
tools/              build + smoke scripts (not shipped in the tarball)

Local development:

node tools/prepare-content.mjs      # copy style markdown into the package
node tools/build-content-manifest.mjs
node tools/smoke.mjs                # host half
node tools/check-client.mjs         # browser half

License

MIT. Third-party style prompts keep their original terms; see content/styles/ for source attribution.

関連プラグイン