dsh-plugin-control-your-video
Identified★ 2DeepSeek Harness Plugin: Making AI Actually Create Videos. A Preset + 13 Method Skills + Forensic Tools; keep your works in your own folders, and turn your experience into your own asset library. A DSH plugin that lets the assistant actually make videos. / MIT
control-your-video
A DeepSeek Harness plugin that lets the assistant actually make videos.
Overview
control-your-video is a DeepSeek Harness (DSH) plugin bundle. It does not ship a timeline editor or a render farm. It ships method: one agent preset that assembles the tools a video project needs, and thirteen skills that tell the assistant how to plan, write, shoot, cut, mix, review and release a video like someone who has done it before.
Your footage, your project files and your accumulated experience stay in your own folders. The plugin never moves them, never locks a format and never uploads anything.
Features
Planning and writing
- Turn a vague feeling ("make it hit harder") into a complete plan: a prose overview, an at-a-glance table, stages, every shot, visual system, sound, rhythm, assets — and a
plan-checkgate that refuses to let production start until it is all there. - Every shot declares where its motion comes from — subject movement, camera move, cut, transition, element animation, full-frame effect, or still — so "the background is shimmering" cannot pass as "this shot moves".
- Strong boundaries between what the assistant decides and what only you can decide — aesthetic direction is yours.
- Rewrite an article into spoken narration; strip AI-flavoured phrasing.
Making the picture
- Decide whether a shot should be filmed, generated, or drawn with code.
- Prompt guides for current video models, with each model's control surface and failure modes.
- Code-rendered frames (HTML / Canvas / three.js) with a deterministic render pipeline and a frame-level acceptance check.
Sound, edit and delivery
- Choose between a human recording, TTS, existing voice-over, or no narration at all — then mix to a measured loudness target.
- Transitions, cutting rhythm, subtitle line-breaking and retiming.
- Review by evidence, not by memory: extract frames, measure black frames, loudness and silence, and keep an index of what was actually looked at.
- A delivery check that separates "technically valid", "rights are cleared" and "I think it looks good" into three different answers.
Everything else
- 1 preset that wires up file access, shell, long-running jobs and on-demand skill loading.
- 13 skills, each under 150 lines, with 49 on-demand references. The body only says when to read which reference, so nothing bloats the context until it is needed.
- 1
evidencetool wrappingffmpeg/ffprobefor frame extraction and measurement, plus aplan-checkaction that gates production on a complete plan. Optional — you can make a video without it, you just cannot prove what you looked at.

Installation
Requirements:
- DeepSeek Harness — the host application. This is a DSH plugin bundle, not a standalone program.
- Node.js 22.3 or later — the preset resolves its skill directory through
process.getBuiltinModule. - FFmpeg and FFprobe (optional) — only the
evidencetool needs them; everything else works without.
Install from a local directory:
- In DSH's plugin manager, choose install from local directory and point it at this package's root directory.
- Restart DSH. Bundle and preset changes only take effect after a restart.
- Open a new session and select the 视频创作 (video creation) preset. Preset changes never apply to sessions that are already open.
- Check the skill catalogue lists 13 names, from
video-intentthroughvideo-skill-authoring.
After installation, editing a skill's markdown takes effect immediately — skills are read from the package directory at runtime. Only changes to cordis.patch.yml require reinstalling and restarting.
Usage
Say what you want in plain language. A brief, a paragraph of copy, a folder of footage, a reference film, or "the part at 0:40 drags" — all valid openings.
The first round normally is not a finished video. You get three things:
- What it thinks you want — your "make it feel expensive" translated into concrete picture, pacing and sound decisions, with the uncertain parts stated as uncertain.
- A plan — format, structure, duration budget, and what each shot needs.
- A short sample — when the project is worth testing on a small piece first.
It stops for you three times by default: after confirming the intent, after the complete plan, and after the sample. Everything between — sourcing, rendering, checking — it runs on its own, and you can interrupt at any point. When it stops, it says what was decided, what is still open, and what you should look at or answer.

Your Asset Library
This is what separates the plugin from a one-shot prompt: finished work leaves something reusable behind.
On your first project the library is empty, and that is correct. It is not a bundled reference collection; it grows out of your own films. When a project ends, the assistant picks out the practices that genuinely generalise and asks whether to record them. Only after you say yes does the first entry exist.
<workspace>/control-your-video/ ← cross-project experience, travels with the workspace
index.md one line per entry; the only way in when reading
preferences.md your aesthetic preferences and hard constraints
methods/ method cards: under which format, what happens, when it stops working
prompts/ recipes measured on this machine, with version and failure conditions
references/ break-downs of reference films: what was borrowed, what was not
templates/ reusable section structures
lessons.md traps hit, and how to recognise them again
Project files — plans, footage, renders, evidence — stay out of the library, in your own project folder. The library only takes what still holds for a different film.
- What is worth recording, and how to judge whether an entry qualifies: the
video-libraryskill. - Field definitions and the single-line
index.mdformat:docs/编写规范.md§1.4.
Can you use someone else's library? Yes. It is a handful of Markdown files. But every entry carries its own applicability and failure conditions, so read those two lines before copying. Sharing your own library is equally fine — those are your files, under whatever licence you choose.
Project Structure
skills/ 13 skill directories, each with SKILL.md and its own references/
tools/ evidence.js (frame extraction, measurement) + plan-check.js (plan gate), both runnable by hand
docs/ 编写规范.md (the five shared contracts), DETAIL.md (plugin page copy)
locale/ plugin-manager strings, Chinese and English
assets/ cover and real stills used by the READMEs
index.js thin DSH tool wrapper around tools/evidence.js
The five shared contracts — shot spec, intent brief, revision brief, library entry, evidence index — are defined in exactly one place, docs/编写规范.md. Skills reference them and never restate them.
Limitations
Read this before trusting anything.
- Tested end to end once. One 5 min 12 s science-explainer video was taken from zero to delivery. The full path has never been run twice with a library to reuse.
- No A/B comparison. There is no evidence that this plugin beats a plain prompt. It has not been measured.
- Not watched or listened to. That run had no audio playback and no continuous playback channel, so every impression was inferred from measurement and extracted frames. Nobody sat and watched it straight through.
- No library reuse ever exercised. A library was produced exactly once (7 categories, 8 files). A second film that reads it has never happened.
- Generation models untested. The prompting skill was loaded and judged "not applicable" on that project; no real generation was run.
- Remotion and Canva untested.
gcuntested. The evidence tool's garbage-collection action has never run on a real project.- Platform AI labelling and fact-checking are out of scope. Both remain the user's call.


Notice on Source Paths
Skills cite provenance in paths like resources/research/…. Those point to the author's research library:
- The material is not distributed with this package and normally does not exist on your machine.
- Full paths are written so that anyone holding the material can trace it, not as a runtime promise.
- When the files are absent, follow the method in the skill body and its
references/— the plugin works without them.
License
MIT.
What ships in this package is method and code written for this plugin, free to use, modify and redistribute. External tools (FFmpeg, Blender, a browser, TTS services) are installed by you and governed by their own terms; being called by this plugin does not relicense them. Your films and your asset library are yours and do not change ownership because you used this plugin.
Comments
Loading…
From the same category
DeepSeek Harness plugin for Reactive Resume: bridges your resumes and job applications into a Harness session over MCP.
★ 41.7k
↓ 63/wk
MIT
Aug 24, 2026
dsh plugin --profile web add dsh-plugin-reactive-resumeby dsh-market
The plugin market inside DeepSeek Harness — browse, search, one-click install · DSH 可视化插件市场
★ 13k
↓ 271k/wk
MIT
TypeScript
Nov 19, 2026
dsh plugin --profile web add dshmarketby kingcharleslzy-ai
Community-maintained dsh-purge fork with host-policy patches, prompt and rule settings, diagnostics, separate update-check and install controls, and local backup/rollback.
★ 10.7k
MIT
JavaScript
Nov 18, 2026
dsh plugin --profile web add dsh-purge-communityby Tencent
Let AI agents use your real, logged-in browser without interrupting your work. CLI + extension for browser automation across any shell-capable AI agent.
★ 10.6k
↓ 17k/wk
MIT
TypeScript
Nov 19, 2026
dsh plugin --profile terminal add @wxg-prc-cpg/browser-skill-dsh-pluginby superdesigndev
OpenRouter for agent tools. Join community here: https://discord.gg/6mQYYfFMAn
★ 8.5k
NOASSERTION
Python
Nov 20, 2026
dsh plugin --profile web add treg-dshby yjh051108
dsh-routing-suite — injector + router-standard kit: install the runtime injector first, then the task-aware reasoning-mode router preset (measured P1-P23).
★ 7k
MIT
JavaScript
Sep 18, 2026
dsh plugin --profile web add @dsh-external/dsh-super-injector