Video Walkthroughs Without the Screen Recorder
How SupportPages produces narrated MP4 walkthroughs from source-grounded mockups, and why regenerating a video is cheaper than editing one.
Time yourself making a ninety-second product video properly. Set up the account state so the screen shows something sensible. Close the notifications. Find a room where nobody will talk. Record, fluff the third sentence, record again. Trim the dead air at the start, add captions because most people watch muted, export, upload, write a description. Forty-five minutes is optimistic, and that is for one video, once, on a UI that will change.
Nobody does that eleven times. So teams make five videos at launch and then stop, and the five slowly become a museum.
Why videos rot faster than prose
A stale paragraph is often still mostly right. "Open Settings, find the notifications section, turn on weekly digests" survives a fair amount of interface drift, because it describes intent and the reader adapts.
A video does not have that property. It shows one specific interface, frame by frame, with a cursor moving to one specific position. Move the button and the video is not approximately wrong, it is visibly wrong - and visibly wrong is worse than absent, because it broadcasts that nobody has looked at this in a year. Every viewer sees the discrepancy in the first four seconds.
So video documentation has a strange economic profile. It is the most effective format for showing a flow and the fastest to become an embarrassment, which is why it is simultaneously the thing everyone says they want and the thing almost nobody maintains.
Generating rather than recording
The walkthrough pipeline reuses the machinery that already exists for illustrated articles, which is the reason it is feasible at all.
Generating an article already requires understanding a flow well enough to write "step three: click Update", and rendering mockups already requires reconstructing the relevant screens from your compiled CSS and design tokens. A walkthrough takes those same mockups, links them into a click path, and drives Puppeteer through it - moving a cursor, typing into fields, transitioning between screens - while recording. Intro and outro cards are composited from your branding.
Where an ElevenLabs key is configured, the same script that drove the recording becomes narration in a synthetic voice, and you get a narrated MP4 plus subtitles plus a text transcript. Without one, the video is silent, which is genuinely fine given how much video is watched muted.
The output is an MP4 with a poster frame, and the whole thing has its own generation lifecycle. Regenerating a walkthrough never touches the article's prose or its review state - deliberately, because the two rot at different rates and you should be able to refresh the video without re-approving the words.
It is a reconstruction, not a recording
It is a reconstruction, with all the limits that implies. The video shows an accurate rendering of your interface's structure with plausible content in it, not a recording of your actual application with your actual data. For "here is how to invite a teammate" that distinction does not matter. For "here is what the dashboard looks like once you have six months of history", it does, and you should record that one yourself.
Narration quality is good and it is not a person who knows your product. It reads the script correctly, stresses the wrong word occasionally, and pronounces unusual product names however the model thinks they should sound. There is a voice picker; there is no pronunciation dictionary yet, which is the fix that would actually help.
Video generation is also the slowest thing we do by a distance - minutes per walkthrough, not seconds. That is fine for a background job and it means "regenerate everything" after a redesign is an overnight operation rather than an instant one.
And the recommendation engine deliberately caps how many articles it suggests for video. Not every article benefits, most benefit less than the effort implies even when the effort is automated, and a help centre where everything is a video is worse than one where the right ten things are.
What changes when recording is free
The interesting shift is not quality. Videos made this way are not better than a careful screencast by a person who knows the product.
The shift is that "which five features deserve a video?" stops being a rationing question. When each one costs a job rather than an afternoon, the answer becomes "the ones that are genuinely easier to show than to tell" - which is a content decision rather than a budget decision, and a much better basis for choosing.
Look at your help centre and count the videos. Then check when the interface in them last matched the product.
About the Author
Simon Moxon has been building SaaS companies since 2009. He's scaled teams to 20+ people, hit £2m ARR, and learned most of the lessons the hard way. Now he's building SupportPages.io - help docs that write themselves when you ship code. He writes about what's changed in SaaS and what hasn't.