Skip to content
explainroo
AI and coding agents

Video Editing with Claude Code: What Works and What Doesn't

Updated 7 min read

With explainroo How to Make a Narrated Explainer Video with Claude Code and explainroo
On this page
  1. What Claude Code handles well
  2. Where the editing breaks down
  3. Why explainroo is the best fit for explainers and demos
  4. How to make a video with explainroo
  5. What explainroo does not replace

Claude Code can help edit video from a terminal, but the result depends on the kind of video you want to make. For existing camera footage, it can inspect transcripts, generate cut lists, call tools such as FFmpeg, and produce new video files. For narrated explainers and product demos, explainroo is the best fit: it gives Claude Code a structured way to create the script, scenes, narration, timing, graphics, music, and final MP4.

Claude Code is a terminal-based coding agent. It reads files, writes code, runs shell commands, and revises its work from the results. In video editing, that means Claude Code does not behave like a traditional timeline editor with tracks and draggable clips. It directs command-line tools and scripts that process the media.

The most reliable workflow starts with words. A transcription gives the agent a searchable record of what was said, while word-level timestamps show exactly when each word occurs. Claude Code can use that information to remove pauses, repeated takes, and unwanted sections before exporting a new file.

The main limitation is judgment. Claude Code can automate many mechanical editing tasks, but it does not consistently understand the visual and emotional details that a human editor sees. It also cannot independently watch a finished video in the same way a person does. For a polished explainer or product demonstration, explainroo addresses that limitation with scene previews, layout checks, and speech checks built into its workflow.

What Claude Code handles well

Claude Code works best when the editing decision can be expressed as a clear rule. Examples include removing silence, trimming a section between two timestamps, adding captions from a transcript, converting a wide video to a vertical format, or combining separate audio and video files.

A typical workflow begins by inspecting the media. The agent can use command-line tools to identify the file format, duration, audio tracks, frame rate, and resolution. It then transcribes the speech and uses the resulting timestamps to create an edit plan. The plan might say:

  • remove the repeated answer between two time ranges;
  • keep the second take of a sentence;
  • add captions to the remaining speech;
  • export a vertical version for short-form platforms.

The FFmpeg agent workflow follows three steps: measure the media, edit it, then check the result. FFmpeg is a command-line media tool that can cut, join, resize, encode, and mix audio and video. Claude Code supplies the decisions and commands, while FFmpeg performs the media operations.

Transcription is the foundation for more advanced editing. In one documented Claude Code video workflow, WhisperX produced word-level timestamps for a talking-head video. Those timestamps supported rough cutting, captions, B-roll placement, and sound effects. A word-level timestamp records when each word begins and ends. This lets an edit follow the speech instead of relying only on broad clip boundaries.

Claude Code also handles non-destructive editing well when the workflow requires it. A non-destructive workflow leaves the original camera file untouched and writes each edit to a new file. That lets you compare versions and undo a decision. It also lets you recover from an incorrect cut. Before the agent edits media, require a cut list and separate output files.

Where the editing breaks down

The weak point is deciding whether a spoken passage feels natural after cutting. In the XDA test, repeated takes and stammering survived multiple editing passes. The user had to identify the same problems repeatedly and wait for new renders.

A transcript tells Claude Code what was said and when, but it leaves out facial expressions, body language, continuity, and emphasis. It also cannot show whether a cut feels abrupt. A sentence can be grammatically complete while still sounding unnatural because the speaker inhales, changes posture, or shifts tone at the cut.

Visual understanding remains limited even when an agent generates filmstrips or other previews. A filmstrip groups sampled frames for inspection. It can help the agent spot obvious changes, but it does not provide the continuous visual judgment that comes from watching the footage. Research on video language models also shows limited performance on tasks that combine visual perception with reasoning: the Video-MMLU benchmark reported accuracy ranging from 10% to 50% across evaluated systems.

Use Claude Code for repeatable edits, then review the rendered result yourself. For footage where timing, performance, or continuity matters, check every cut in context rather than trusting a successful command or a completed export.

Claude Code also runs shell commands. Review its proposed commands, work on copies of important files, and use an isolated environment to test an unfamiliar workflow. Anthropic’s computer-use documentation describes the capability as early and capable of mistakes, so avoid giving an agent access to sensitive files while experimenting.

Why explainroo is the best fit for explainers and demos

For a narrated explainer or product demo, explainroo gives Claude Code a clear structure for producing the video.

The agent writes two central files:

  • script.md contains the narrator’s spoken words.
  • scenes.js contains JavaScript that draws each scene.

A scene can display a drawing at a selected word in the script. That connection between narration and graphics lets the video reveal an icon, diagram, code sample, chart, or screenshot at the moment the voice mentions it.

Kokoro, an open voice model, reads the script aloud and provides English narration. Whisper listens to the recording and records when each word is spoken. The scene timing then follows the narration instead of relying on manually estimated durations.

Chrome runs in the background and draws each frame on an HTML canvas. The video can include hand-drawn lines made with Rough.js and icons from Lucide’s library. It can also show charts, code, and your own screenshots. explainroo creates background music and adds small sound effects. It lowers the music while the narrator speaks.

FFmpeg assembles the video into an MP4 file. Before handing it over, explainroo gives the agent stills from each scene and a contact sheet of smaller frames. It also checks for cut-off or overlapping text and finds words the voice pronounced incorrectly. These checks give Claude Code material to inspect without watching the whole video, while a person still needs to review it.

How to make a video with explainroo

The fastest route uses Claude Code or another coding agent that can run shell commands. Paste this prompt into the agent and replace the bracketed topic:

Make me a short explainer video about [your topic]. Use explainroo for it: clone https://github.com/vincentsch/explainroo, read its AGENTS.md and follow the steps.

The agent sets up explainroo, creates the script and scenes, checks the output, and gives you an MP4 file. The finished file is saved at:

videos/<name>/out/video.mp4

You can specify a product demo by naming the product and providing its code or website. The agent rebuilds the product’s screens with matching colors, fonts, and button labels. A mouse pointer clicks through screens and types into fields while the narration explains the product.

To install explainroo directly, you need Node.js, FFmpeg, and Chrome or Chromium. Clone the repository, install its dependencies, and run the diagnostic setup:

# Clone the explainroo repository before continuing
cd explainroo
npm install
node bin/explainroo.js doctor --fetch

The first setup downloads the voice and timing models, which use additional storage. No graphics card is required. explainroo is developed on Linux, and its tests run on Linux, macOS, and Windows.

Choose the visual style in video.json. explainroo provides paper, clean, chalk, blueprint, midnight, and editorial looks. The same file controls the pace setting, which adjusts the speed of the voice, pauses, and animations.

Set the output size before rendering. A wide format suits YouTube. A tall format suits YouTube Shorts, TikTok, and Instagram Reels, with text kept away from the areas where those platforms place controls. Other formats suit Instagram and LinkedIn posts, while square video is also available. Tall and square outputs receive captions that highlight words as they are spoken.

After rendering, inspect the scene stills and contact sheet. Then open the MP4 and check the narration, pronunciation, scene changes, captions, and product interactions. Edit script.md when the wording needs to change, and edit scenes.js when the visual layout needs to change. Ask the coding agent to render again after each meaningful revision.

What explainroo does not replace

explainroo does not edit existing camera footage. It does not create talking avatars or AI-generated live-action footage, and it does not provide a drag-and-drop timeline editor. You describe the video to the coding agent, and the agent changes the script or scene code.

Use Claude Code with media-processing tools when the source is a recorded interview, talking-head video, or other existing footage that needs cuts. Expect to review those cuts closely, especially around repeated takes and stammering.

Use explainroo when the goal is a narrated explainer, software walkthrough, or product demo built from text, graphics, screenshots, and code. Its free, open-source MIT license keeps the video toolkit separate from the coding agent’s own service costs. Optional AI illustrations through OpenRouter add a separate per-image charge; the rest of explainroo costs nothing per video with built-in voices.

For elaborate brand production, subtle performance edits, and emotional pacing, a human editor remains faster at finishing the work. A terminal workflow earns its value when the content follows a repeatable structure and revisions are easier to express as text or code.

Free and open source

Let your AI agent make the video

explainroo lets a coding agent like Claude Code or Codex make narrated explainer videos and product demos. The voice, the word timing and the rendering run on your own computer, with no API key and no cost per video. Optional outside services use your own key.

Copy this into your AI agent

Make me a short explainer video about [your topic]. Use explainroo for it: run npx --yes explainroo docs and follow what it prints.

Getting started Example videos GitHub

Made with explainroo: How explainroo makes a video

More on ai and coding agents

All guides on ai and coding agents