Preface¶

When producing AI short films using multiple models, the common practice is: the agent writes the story and prompts, the image model generates the first and last frames, and the video model generates shots billed by the second. Video regeneration is charged by the second, while prompt editing costs almost nothing—therefore, you need to filter out text and visual issues before moving to image generation or video generation. DeepSeek Harness (DSH) places the agent conversation within a Web UI, but storyboards, prompt libraries, and shot pipelines are often scattered across notes or spreadsheets. dsh-video is a community-maintained DSH plugin that adds a Video Studio panel to the left sidebar, bringing storyboards, prompt assembly, and the shot pipeline into the same session environment alongside the agent conversation.

What is this¶

dsh-video (repository xuboboo/dsh-video) is published by maintainer xuboboo and categorized as a Workflow in the SkillHub directory. The description states: an AI video production studio panel for DeepSeek Harness—storyboard, prompts, and shot pipeline. The plugin is a pure JavaScript Cordis bundle, requiring no build steps and no additional runtime dependencies; the license is MIT. The README footer labels the release as v1.0.0, while the version field in package.json inside the repository is 2.0.0; please refer to your local checkout.

Core Features¶

The following capabilities come from the repository README, summarized by panel structure.

Sidebar and Three Tabs¶

The plugin registers the Video Studio in the full-screen panel on the left side of Harness, alongside Plugins and Scheduled tasks, containing three tabs: Studio, Prompt library, and Failure modes.

Studio: From Project Init to Shot Pipeline¶

• Director brief: The complete prompt for starting multi-model work, including role division (you / director / painter / animator), Codex CLI call method, Seedance API endpoint, as well as model IDs and 720p reference prices (e.g., doubao-seedance-2-5-260628 ≈ ¥1.5/s, doubao-seedance-2-0-mini-260615 ≈ ¥0.5/s) and seven workflow rules.
• Pre-flight check: Checks the painter tool, ffmpeg, Python, ARK_API_KEY, etc., per project and generates a copyable report; generation does not proceed until the environment is ready.
• Story & style card: One-line information, hook, story structure (episodic vs. continuous), art style, character sheet, palette, music, and other fields, with relevant rules printed under each field.
• Storyboard shots: Each shot includes plot, active action, moving objects, continuity with the next shot, shot size/camera/motion, composition (including 38%/62% rule), estimated seconds, beat-by-beat, frame description for the image model, sound, etc.
• Seven-segment prompt assembly: One-click assembly of Chinese and English prompts structured as [Style][Shot][Composition & Color][Character][Beat-by-beat][Sound][Constraint].
• Frame checklist: 8 items per shot for start/end frame verification (props, pedestrians, pose repetition, headroom, wind consistency, phone readability, clutter, overlay drift, etc.).
• Cost ledger: Records the model, seconds, retained/failed, files, and failure reasons for each submission. Calculates pricing by model ID and compares it with the project budget cap.
• Review decision: Fields for cut / fix code / regenerate decisions, and prints priority rules from the article.

Prompt library and Failure modes¶

• Prompt library: 10 sets of Chinese and 10 sets of English copy-paste pipeline prompts (director brief, story development, story audit, storyboard rules, start/end frame audit, pre-submission audit, post-production, final review, four-line team mode, etc.).
• Failure modes: 20 lines of typical video model failure modes and corresponding prompt adjustments for pre-submission self-check.

Language, Storage, and Optional Cloud Sync¶

• Bilingual UI: One-click toggle between Chinese and English, defaulting to the browser language. The interface uses Harness theme tokens and changes with light/dark themes.
• Local-first: Data is stored in the browser localStorage, with no account and no server-side telemetry; v1 projects can be migrated automatically.
• Cloud sync (optional): Configurable endpoint to implement the dsh-video sync protocol. Push uploads, Pull & merge merges by project ID and updatedAt (last write wins, snapshot retained). The repository server/ provides a zero-dependency self-hosted reference implementation. ARK_API_KEY does not enter the browser; the sync token is held in memory by default.

The workflow concept mentioned in the README originates from the article “Claude Code/Codex + Seedance AI video: from an idea to a finished film”: a director agent that iterates on text first and finally runs the expensive video generation.

Installation and Enablement¶

Prerequisites: You must have a working DeepSeek Harness (Desktop or dsh web) installed locally. Supports Windows, macOS, and Linux.

Non-Desktop Environment (CLI + Web profile)¶

The installation and startup steps given in the README are as follows:

# First time: create profile from built-in web template
dsh video --from-default-profile web

# Install plugin from local checkout directory (replace path with your dsh-video directory)
dsh plugin --profile video add /path/to/dsh-video

# Start Web UI
dsh video --no-open --host 127.0.0.1 --port 10891

You can also install via a Git URL in environments that support it by accepting the URL after publishing or forking the repository (the README does not hardcode specific GitHub shortcuts; do not manually concatenate install strings not appearing in the documentation).

Desktop Environment¶

In the Harness plugin manager: Plugins → Add → local directory, select the plugin directory, and restart the application to load the new client module.

The plugin runs under DSH process permissions; please review the source code and MIT license yourself before installing, and confirm the path source is trusted.

Typical Usage¶

The usage sequence in the README can be summarized as:

1. Click the dsh-video icon in the full-screen panel list on the left.
2. New project: Name the project, set the aspect ratio, and target duration.
3. Add shot: Describe the scene and camera motion, and fill in the frame prompt for the image model and the video prompt for the video model.
4. Walk through the pipeline from top to bottom: fix text first, then generate frames and check, and finally generate video; update the status badge after completing each shot.
5. Copy the Director brief to a new Harness session to start multi-model collaboration with the agent.

If multi-device synchronization is needed, you can first self-host the sync service (as shown in the README):

SYNC_TOKEN=my-secret node server/server.mjs   # Default 127.0.0.1:8787

Fill in the endpoint (e.g., http://127.0.0.1:8787) and token in the plugin Cloud sync, perform Push locally; on the other device, perform Pull & merge with the same endpoint.

Suitable Scenarios and Notes¶

Who is this for: Teams or individuals already using agents in DSH to write films, paired with image models (such as Codex) and video models (such as Seedance); people who need to place storyboard fields, prompt templates, cost ledgers, and failure mode comparison tables in the UI rather than scattered in documents.

Notes:

• Video is billed by the second. The plugin provides a cost ledger and a workflow constraint of “text first, then images, then video,” but it does not replace your management of API keys and credit limits; the README explicitly states rules like “never top up credits” are written in the director brief.
• Data is stored locally in the browser by default; clearing the cache or changing browsers will lose projects unless cloud sync is enabled or you export them manually.
• The community directory SkillHub plugin page and the GitHub repository are independent of DeepSeek / Huansuan and belong to third-party workflow plugins in the “everything is a plugin” ecosystem; GitHub stars may be 0 when scraped by the directory, which does not represent a quality judgment.

Closing¶

dsh-video brings the director brief, story and storyboard fields, seven-segment prompt assembly, start/end frame checklists, cost ledgers, and failure mode tables into the Harness sidebar, used side-by-side with agent conversations, highlighting the production discipline of “fixing errors on text and visuals first, then burning video seconds.” For further information, please check the SkillHub directory page and GitHub repository (including the WORKFLOW Chinese documentation).