Edit video clips with AI instructions using a non-linear timeline editor powered by ffmpeg. Edit videos by describing what you want. Video agencies spending $50/hr on editors can automate 60% of cuts, trims, and overlays.

This is the whole blueprint, nothing held back: the model, the tools, the exact prompt, the source on disk, and the setup. Everything you need to run Video Editor yourself is on this page, and it deploys in an afternoon.

See it running

Agents like Video Editor run behind a studio with a live preview of what it makes. Here is that pattern, live: every action stays visible and reviewable. Click into it.

crescendo.51ultron.com/kits/builder-studio

What it does

Video Editor handles the work end to end:

Edit video clips through natural language instructions
Non-linear timeline with multi-track video and audio
Trim, move, overlay, adjust volume, and add text overlays
Preview renders and full exports via ffmpeg

The anatomy

An agent is a composition, not a prompt: a model, the tools it is allowed to call, the integrations that give those tools reach, and the rules that keep it honest. Here is Video Editor on one card.

Video Editorblueprint

Model: claude-sonnet-4-6
Integrations: ffmpeg
Guardrails: never guess asset duration. probe first if unknown.; video tracks stack bottom-to-top. higher tracks overlay lower ones

The instructions

Every run is governed by this prompt. It is short on purpose: enough to make Video Editor reliable, not so much that it can misread itself. Copy it as a starting point for any agent you build.

Video Editor

You are a non-linear video editor assistant. You operate a timeline editor that runs ffmpeg in the sandbox. Every user message is prefixed with: [[[SYSTEM NOTE: PROJECT: <JSON> ]]] That JSON has { tracks, clips, assets, output }. Treat it as source of truth. WORKFLOW: 1. Read the SYSTEM NOTE to understand the current timeline. 2. If new URL provided, call probe_asset to learn duration/dimensions. 3. Plan edits as small ops (add_clip, move_clip, trim_clip, set_volume, set_text_overlay). Emit all in ONE update_timeline call. 4. Only call render_project when user asks to render/export. RULES: - Never guess asset duration. Probe first if unknown. - Video tracks stack bottom-to-top. Higher tracks overlay lower ones. - Keep timelines under 60s unless asked otherwise. - Be concise. One short sentence + one tool call.

The source, on disk

Three files carry the whole agent: the definition with its tools, the environment it expects, and the setup. Walk them.

video-editor/video-editor.ts0.5 KB

1import { agent } from "@agent-sdk"

2import { videoEditorTools } from "./lib/tools"

4export default agent({

5 model: "claude-sonnet-4-6",

6 runtime: "claude-code",

7 permissionMode: "bypassPermissions",

8 maxTurns: 25,

9 systemPrompt: `...`, // see System Prompt above

10 tools: videoEditorTools,

11 onError: async ({ error }) => {

12 console.error("[video-editor] error:", error)

13 },

14 onFinish: async ({ cost, duration, turns }) => {

15 console.log(`Done: ${turns} turns, ${duration}ms, $${cost.toFixed(4)}`)

16 },

17})

The business case

If you run an agency, Video Editor is also a product. It sells into content creators, marketing teams, video agencies, and the numbers work at very small scale.

$4,000

setup you can charge

$600/mo

monthly retainer

60%

less manual work

Ship it

The whole install is a handful of commands and the keys from the env file. Run the list, then point Video Editor at real work.

Deploy checklist

0 of 5

That is the entire blueprint: one model, a clean toolset, and a prompt. Built by hand it costs you an afternoon. Inside Ultron it costs you a sentence, because you can build this exact agent in the workforce and watch it run before you commit to anything.

Build Video Editor in Ultron

Describe what you want

Start from this blueprint or your own words. Ultron reads the intent and picks the tools.

It builds the agent

The model, the tools, the prompt, the console. Assembled in front of you, nothing to wire.

Watch it run

Test it live in the chat, then put it to work. You review what it does, it does the rest.