GuideGen
Do the task once. Get the guide, the narrated video, the shareable link, and a click path an AI can read.
Live and free. Runs on your own machine.
What it does
Records a browser workflow while you work and writes the documentation for it — annotated screenshots, numbered steps, and five ways out.
A narrated video, rendered on your laptop
One slide per step at 1080p with a voice reading it aloud. The speech model runs in the browser tab, so the text of an internal process never reaches a speech service — and there is no per-video cost to pass on.
Capture the last 2 minutes
Nobody knows in advance that they are about to want a guide. Switch it on for one site and the guide can be built out of work already finished. Off until enabled, per site, deleted after 7 days.
A click path for an assistant
One click copies every step as Markdown with the URL it happened on and the requests each step fired. Explaining an internal tool to a model usually takes longer than doing the task.
Publish a link, or export a file
Web page, Markdown, PDF, slides or video — every one built inside the browser. Publish one guide deliberately and the link updates in place.
How it got built
The problem
Documentation is written by the person with the least time to write it, and it rots the week after. The actual job is screenshot, crop, label, repeat — an afternoon of work that produces a document nobody reads and nobody updates.
The constraint that shaped it
It had to be free at any volume, which is a design constraint rather than a pricing decision. Every competitor renders on their own servers and pays per GPU-minute, which is why they all gate hard. That ruled out a server for the expensive parts and forced the whole pipeline into the tab.
Decisions
- Read the page instead of asking a model
Step text comes from the accessibility label and role of the control that was actually clicked — so it says Click “Team settings” rather than a model’s guess. It is exact, it is free, and it works offline. The screenshot is cropped to the part of the page that matters instead of shrinking a window until nobody can read it.
- Rejected the browser’s built-in model, on purpose
Chrome ships a free on-device LLM and it was the obvious choice. I tested it and turned it down: shaky prose sitting beside counted numbers makes the numbers look invented. A documentation tool that is occasionally eloquent and occasionally wrong is worse than one that is always literal.
- Stop asking people to predict the future
The first version needed you to press record before doing the thing. That is the one moment nobody gets right. A rolling two-minute buffer moved the decision to after the work, which is when somebody actually asks how you did it.
Where it landed
Published and public on the Chrome Web Store. It took five uploads and four rejections, three of them for one policy — remotely hosted code in a Manifest V3 item — and each rejection had a different source.
What it taught me
The fix that mattered was not a cleverer patch. The automated check had been pointed at the file that caused the previous rejection, so it could not possibly find the next one. It now sweeps every text file in the built package. A check written to catch the last failure is a check that will miss the next one.
What I’d do next
Guided mode: open a guide and have it point at the real control on the real page, advancing as the person clicks. It points, it never clicks — which keeps it out of automation, and is the reason this is available here and not to tools that took the replay route.