--- title: "Turn any website into a branded promo video — from just a URL" date: 2026-07-07T11:50:41.162+00:00 url: https://hurricane.works/blog/turn-any-website-into-a-promo-video description: "We built a tool that watches a web app the way a person would, then produces a narrated, on-brand promo video from nothing but a URL — and the whole thing was built, and debugged, by an AI coding agent that could see its own video output." author: "Hurricane Works" categories: ["Video"] --- # Turn any website into a branded promo video — from just a URL We built a tool that turns any web app — or a public website — into a set of branded promo videos, from nothing but a URL. Point it at a site and it produces a full narrated walkthrough and a short punchy teaser, each in a sound-on cut and a silent captioned cut for autoplay. ![YouTube video](https://www.youtube.com/watch?v=KuNVyxB99m8) On brand, 1080p, with a virtual camera that glides and zooms into whatever the narrator is describing and real numbers counting up on screen. The sort of video an agency charges four figures and a week for — in a couple of minutes, for a few pennies of API calls. And the whole thing was built, end to end, by an AI coding agent. More on that below — because it's the part that changes how you might work, not just what we shipped. ## Why you'd want this Showing beats telling. The fastest way to make someone *get* a piece of software is to show it moving — which is why every landing page wants a demo video and every launch wants a walkthrough. But a good product video is a whole project: record the screen, write a script, get a voiceover, edit it together, add captions, make it on-brand. Most teams make exactly one, for the homepage, and never again. The bottleneck was never "should we have a video" — it's the time and cost per video. Remove those and the maths flips: you make one for every feature, every use case, every client, and you keep them current instead of letting them rot the moment the product changes. That's who this is for. Founders launching a product or a feature. Anyone with a page that would convert better with a demo. And especially **consultants and agencies** — people who'd love to hand every client a polished promo video but could never justify the hours to do it bespoke, one client at a time. ## Why it's clever: it *learns* the site first The magic is a step we call **learn**. You don't script anything. The tool opens your site and looks at it the way a person would: it reads your real brand colour and logo straight off the screens, works out what each page is for, writes a walkthrough script and a short teaser, drafts the on-screen card copy, and even writes an SEO-optimised YouTube title and description ready to paste. Then it films. It screenshots each screen, and for the walkthrough a virtual camera pans and zooms into exactly the heading or figure the narrator is naming — with the real numbers ticking up from zero — while a natural voice reads the script with proper pauses. No editing timeline, no After Effects, no stock-music faff. Crucially, the draft is just a file you can read and tweak before it renders. And re-rendering is essentially free, so when your product changes you regenerate the video in minutes rather than booking another shoot. ## How you'd actually use it One command: give it a URL and pick a voice. - **Your own app?** Run it against your local or staging site. - **A client's product?** Point it at their live public site. You need no code and no access — it handles the real-world mess of production sites (dismissing cookie banners, triggering lazy-loaded images, following the site's internal links). Out comes four video cuts — a full walkthrough for sales calls, onboarding and docs, and a short sizzle for the landing page and socials, each in a sound-on and a muted-captioned version (landing pages autoplay silently, so the captions carry the message). Plus the ready-to-paste YouTube copy. It picks up the brand automatically, so it looks like *them*, not like a template. For a consultant, that's a genuinely new service you can offer at near-zero marginal cost. For a founder, it's the difference between "we should really make a demo" and having one before your coffee's cold. ## The benefits, plainly - **Speed and cost collapse.** Minutes and pennies instead of days and a production budget — so you make videos you'd otherwise never have made. - **Always on brand.** It reads your colours, logo and font off your own site. - **Always current.** Product changed? Re-render. The video never goes stale. - **Two audiences, one run.** A long walkthrough to explain, a short sizzle to hook — from the same pass. - **Repeatable as a service.** One tool, any site, any client. ## The part that'll interest anyone who codes: an agent built it We didn't build this tool by hand. An AI coding agent (Claude Code) built it end to end across one working session — and, more surprisingly, **verified its own video output.** It didn't just write code and hope. It ran the pipeline, rendered the videos, and *looked at the results*: reading rendered frames to check the brand colour and layout were right, and diffing frame hashes to prove a bug rather than guess at one. When a tester said a teaser "looked broken", the agent pulled three frames a couple of seconds apart, hashed them, saw the hashes were identical, concluded the camera was frozen, traced it to a locator-matching bug that only fired on one anchor format, fixed it, and re-rendered — before a human looked again. The division of labour is the interesting bit. The agent did the building *and* the checking; the human did the steering — scoping the phases, catching that a voiceover "read like a bulleted list being read out", choosing the narration voice, deciding which pages to feature on a big client site. It built the machine and drove it; we pointed it and judged the output. The lesson for anyone using these agents: the real unlock isn't that they write code — it's that, given a way to *observe* the result (a screenshot, a rendered frame, a number), they'll debug against reality like an engineer instead of emitting plausible-looking code. Close that loop — hand the agent eyes on its own output — and it stops guessing and starts checking. ## Under the hood, briefly The shape is worth stealing even if you never make a video: a **learn pass** where a vision model drafts a reviewable artefact (brand + storyboard + copy), then a **deterministic render**. Creative judgement goes to the model; repeatable assembly stays in code. Two things the agent learned wiring it up: **Let the vision model make the perceptual calls.** The first brand-detection used CSS heuristics — and picked a near-black button as the "accent" and a run-on navigation string as the logo. The model, already looking at the screenshots, nails both. Don't leave a heuristic doing a job a vision model does better right next to it. **When two generated things must line up, make it a contract.** The camera and the voiceover were generated separately and drifted — the camera on a price while the voice said "pick a style". The fix was a single prompt rule: *the camera zooms to the anchors in the order you list them, so write the voiceover to walk those same anchors in that same order, one clause each.* Now they move in lockstep. ## The takeaway Video is the best way to show software, and the only reason we don't have more of it is cost. Collapse that cost and the behaviour changes — you make a promo for every product, feature and client, and you keep them fresh. But the deeper lesson is about how it got built: an agent that could *see* its own rendered output debugged it like an engineer. If you build AI tools, both ideas transfer — "let a vision model draft a reviewable plan, then render it deterministically", and "give the agent a way to observe the result and it will check its own work." That's increasingly the shape of how this kind of software gets made.