Diffusion Studio Skills

Two skills connected to a local video editor

Use watch to find evidence in footage, then editor to turn approved moments into an editable, inspectable timeline.

By diffusionstudio

Diffusion Studio Skills project image

Hosts

  • Claude Code
  • Codex
  • Assistants with access to the local CLI or MCP
GitHub stars
146
Version
1.0.0
Checked

Find one specific action first

Choose a local video you can replay yourself and ask a concrete visual question. Establish that the app, paths and frame tools work before adding transcription or editing, so each step produces something you can inspect.

Pick a task and get started

Configure both the skills and desktop app, then choose a task. Plugin entry names are diffusion:watch and diffusion:editor; standalone installations use watch and editor. Replace the path, scene-id and input placeholders with real values.

Choose the next step

Start with watch on a local video: survey the footage, then inspect the relevant source frames.

  1. Install the two skills

    The author’s general installer uses Node.js/npm. Select your assistant and the watch/editor skills, then reload them. This installs guidance; install the desktop app separately for your platform. The command fetches the version available when run; this page reviews 7e2959c.

    npx skills add diffusionstudio/skills
    Read the author’s setup guide

    Connect the skill to the app and first source file

    Install the desktop app using the platform notes below. If it is already installed but the command is missing, use Settings → CLI → Install and open a new terminal. The two commands below print the command list and start the app in the background; open without a directory creates no project. Give the assistant the absolute video path and a new analysis directory, then have it read the installed docs/skills/watch.md. On macOS the docs are usually at /Applications/Diffusion Studio.app/Contents/Resources/docs; on Windows they are at %LOCALAPPDATA%\DiffusionStudio\app-<version>\resources\docs. Send “Find an action” next.

    diffusion --help
    diffusion open --background
    Installation and connection reference
  2. Give your AI this first task

    The four tasks and workflow below are editorial examples of intended deliverables and checks. We have not run the app, processed footage or generated a video.

    Use Diffusion Studio’s watch skill. The video is at [replace with its absolute local path]. Find the first time the presenter turns the product over to show its back. Read the watch guide shipped with the installed app. Probe the duration and video track, survey it with filmstrip, then narrow the search and use grab at explicit times before, during and after the action. Save images in [replace with a new absolute local analysis directory]. Use local probing and frame extraction only: no transcribe, listen or generation calls, and no video edits. Return source-time ranges, descriptions, supporting image paths and uncertainties. If the action is not fully covered, refine the search rather than treating a coarse thumbnail as an exact boundary.

  3. Check what you receive

    • An action log with source-time ranges, descriptions, supporting frames and the actual output paths.
    • Inspect the before/during/after frames and play the candidate interval in the original video. Check that it is the requested action and that an earlier occurrence was not missed.
    Read the author’s notes for this task
More about this skill

Diffusion Studio Skills gives an assistant two ways to work. watch answers what happens in footage and when; editor arranges cuts, captions, graphics and sound. The plugin contains Markdown guidance. The Diffusion Studio desktop app on the same machine performs the actual probing, frame extraction, composition and export.

The version relationship matters: a skill first directs the assistant to documentation inside the installed app, then uses its bundled diffusion command. dapi remains an alias. Projects are SolidJS JSX files; saving recompiles the composition and updates the canvas. A useful delivery can therefore be an editable project, with video export performed when explicitly requested.

Understanding footage does not require calling every model. For an action, inspect the structure and grab relevant frames. For a quotation, use an existing transcript or authorize cloud transcription. For an edit, approve the content and intent before building the main footage and captions. Image, video and audio generation are optional additions.

Sources and verification

Sources checked

Fixed revisions: skills 7e2959c and editor af570c1. We read 52 documents, configurations and key implementations in full and checked commands statically. No installation, model call or media processing was performed.