Demo workflow
How I create narrated product demos with coding agents
The workflow behind my LaserBurn AI tutorials: inspect the real app, write the narration, synchronize actions, and review the export.
Randall Morgan · · 5 min read
Start with one useful result
A demo needs a clear outcome before it needs a recording. For the LaserBurn AI box-generator tutorial, that outcome is a sliding-lid organizer the viewer can customize and export. I identify the controls needed for that result and leave unrelated features out.
The coding agent inspects the actual application before scripting. Labels, option values, export behavior, and loading states come from a walkthrough. If the live application differs from the source, the tutorial follows what the viewer can actually use.
Make narration the timing reference
My Demo Agent runner executes a director manifest: scenes, narration beats, browser actions, and checks. The agent writes that manifest; the runner does not independently decide what a good tutorial should teach.
ElevenLabs speech timestamps let the runner bind a click or cursor movement to an exact phrase. Unpredictable loading waits sit outside those timed actions. The next beat starts once the result is ready, so the exported tutorial does not spend its narration explaining a result that has not appeared.
Keep scenes recoverable
Each scene recreates its own starting state. A failed export scene can be repaired without recording the entire video again. I keep narration, the manifest, screenshots, captions, and output together so a correction is traceable.
Retries need judgment. Repeating a free browser action is different from repeating a paid generation. The manifest permits bounded retries only when the action is safe to repeat; otherwise the failure becomes something to inspect.
Review what the viewer receives
A playable MP4 proves that encoding worked. It does not prove the tutorial is accurate or readable. I inspect the actual export for small text, cursor timing, useful result holds, narration accuracy, captions, and accidental exposure of account information.
- Use a demo account when a walkthrough needs authentication.
- Hold the result long enough for the viewer to understand it.
- Check the promised download or export actually completed.
- Keep provider keys and saved login state out of shared examples.
What the workflow automates
The repeatable parts are narration generation, timed capture, captions, assembly, and repair of individual scenes. Choosing the lesson and judging its quality still require review. The sample on the Demo Video Toolkit page was produced with this workflow against the live LaserBurn AI box generator. The commercial toolkit is being packaged; the sample is a demonstration of the existing process, not a claim of a finished customer release.