Technical proof
CLI, capture and masking, rendering, recorded narration, the speech sidecar, a benchmark harness and the Claude Code skill.
walkthrough
Tutorial videosWalkthrough opens your web app in Chrome, follows the steps you describe and turns them into a narrated tutorial video, with the controls highlighted and your own voice on the soundtrack.
Each stage saves its output before the next one starts. Change one line of narration and only that clip and its scene are rebuilt. A failed run never overwrites a good export.
Say who the tutorial is for and what they'll achieve, then list the steps in flow.yaml along with what should be on screen after each one.walkthrough init · walkthrough plan
Chrome performs each step on your local app and screenshots the result. Sensitive fields are blacked out before anything is saved.walkthrough capture
Check every frame, its highlight and its masks on one page. Move a highlight or change a crop if the suggestion is off.walkthrough storyboard · walkthrough annotate
Narration is drafted from the labels that were actually on screen. Nothing can claim success before the outcome was captured.walkthrough script draft
Import your own recordings, which get caption timing automatically, or use a local clone of your voice.walkthrough narrate
FFmpeg assembles zooms, callouts, title cards, crossfades and captions into a 1080p video. Preview renders are watermarked.walkthrough render --preview
Approve the storyboard, script and visuals. Edit anything afterwards and that approval is cleared until you approve again.walkthrough approve · walkthrough export
A tutorial that shows the wrong button does more harm than no tutorial. When the app doesn't match the plan, capture pauses and tells you why.
$ walkthrough capture dealer.walkthrough PAUSED [LOCATOR_NOT_FOUND] Step "save": no visible button "Set password" on http://localhost:3000/invite → The interface may have changed. Similar labels on the page: "Create password". Update the step's target in flow.yaml.
Install the CLI, start your app locally with demo data, and make your first project. walkthrough status tells you the next step at any point.
You'll need:
brew install ffmpeg# Install git clone https://github.com/exiscosoft/walkthrough.git cd walkthrough pnpm install pnpm build && pnpm link --global # Make a project, then describe the steps in flow.yaml walkthrough init dealer.walkthrough \ --name "Set up your dealer account" \ --url http://localhost:3000/invite walkthrough plan dealer.walkthrough walkthrough capture dealer.walkthrough # Narrate with your own recordings and render a preview walkthrough script draft dealer.walkthrough walkthrough voice use dealer.walkthrough --engine import walkthrough narrate dealer.walkthrough --from-dir recordings/ walkthrough render dealer.walkthrough --preview
The repository includes a Claude Code skill. Claude plans the flow, runs the capture, writes the narration from what was captured and walks you through review.
Make a walkthrough video showing how a dealer accepts an invitation and sets a password.
Copy skill/walkthrough into your app's .claude/skills/, or link it into ~/.claude/skills/ to use it in every repo. Then ask for a video or run /walkthrough.
Claude stops for the decisions that are yours: which data-changing steps to allow, checking the masks, and the three approvals. Masked screenshots and page snapshots are sent to your model provider when Claude reads them. If you'd rather avoid that, write the flow and script yourself.
Set up the skillWalkthrough is an early technical proof. The full pipeline works and is tested end to end.
CLI, capture and masking, rendering, recorded narration, the speech sidecar, a benchmark harness and the Claude Code skill.
An interactive storyboard editor, brand presets and a privacy review screen, after a speech benchmark and a proof run on a real app.
Three to five marketers making ten or more tutorials. After that: continuous recording, multilingual narration and recapture when the UI changes.
About voice cloning: local voice cloning hasn't passed its benchmark on 8 GB machines yet, so importing your own recordings is the reliable way to narrate for now. Some speech model weights, such as F5-TTS, are licensed for non-commercial use only.