Loading
What a viewer knows, written down.
An API. Give it a YouTube URL. It returns what was on screen, what was said, how it was said, and what was pointed at.
Free. No account. The first three minutes of any public YouTube video.
A proper guide to Fable 5 · Theo · 43:15 · youtube:8GRmLR__OGQ
Talking head · full frame
03:05.2
04:04.0
document2 rows
- 03:05.2seenTheo full-frame at a wooden desk, black crewneck, mic centre; spreads both hands wide, fingers splayed, as he lands “so annoying”.
- 03:05.2THEO(exasperated, softer) the billing side, and trying to get that linked to the user in a way that makes sense. The amount of weird, random, transient billing failures we get across services is so annoying.
Fig. A.1 — Left, the video. Right, what was seen, said, and pointed at, in sync.
A transcript is not the video.
A transcript keeps the words and drops the rest. The slide behind the speaker. The line of code on screen. The URL in the address bar. What “this” meant when someone pointed at it. Which word was stressed, and where the pause fell. On an 8-minute podcast, a language model answered 62.9% of 31 verified questions from the transcript alone. From the Screensay document, 96.8%.
The transcript
- 03:05the billing side, and trying to get that linked to the user in a way that makes sense. The amount of weird, random, transient billing failures we get across services is so annoying.
- 03:13Well, it was until I started using today's sponsor more, Clerk. Because not only do they have the best auth platform that is the easiest to integrate across everything, be it web,
- 03:21mobile, or anything else. We're actually working with them right now in order to get the T3 Code Electron desktop app auth for our new remote stuff that I'm really excited about.
- 03:29exciting to me is the subscription platform that's now built in. Having tried pretty much everything like this, I can confidently say Clerk is the easiest and has the
- 03:37best abstraction. Instead of all the obnoxious effort trying to set up a staging and production environment for Stripe with their own separate keys, their own separate product IDs, their own
- 03:45separate everything, that you now have to mirror in your real database just to make sure it works, you just get to configure it in the clerk dashboard. Once a user
- 03:54pays, that payment information is attached to the user itself, where it belongs. I don't care what subscriptions exist, I care if this user is subscribed,
Everything it left out
- 03:05.2seenTheo full-frame at a wooden desk, black crewneck, mic centre; spreads both hands wide, fingers splayed, as he lands “so annoying”.
- 03:10.0doesgesture: hands open wide — both hands flung wide, fingers splayed, on “so annoying”
- 03:10.0momentfunny: the physical punchline of the complaint that tees up the ad
- 03:13.0seenHard cut to a browser filling the frame: Clerk homepage hero; Theo in PiP top-right looks at camera.
- 03:13.0text“clerk” (logo) · “More than authentication,” “Complete User Management” (heading) · “Start building for free” (button)
- 03:14.8points at“today's sponsor” = “clerk” (screen match)
- 03:23.0momentdisclosure: a stated business relationship beyond the sponsorship; loudest peak of the read on “really excited”
- 03:29.0seenClerk Billing page: hero “Subscription billing, without the headache”; Theo in PiP looks down at his screen.
- 03:30.6points at“the subscription platform that's now built in” = “Subscription billing, without the headache” (screen match)
- 03:32.9momentemphatic: his slowest sentence so far — the endorsement is the one he slows down for
- 03:39.0seenScrolled to “Define and manage plans directly in Clerk”: Starter $9/mo and Pro $19/mo cards; a checkout modal reads “Free trial successfully started!” then clears.
- 03:52.0points at“the clerk dashboard” — not shown; a marketing diagram is up, not the dashboard (unresolved)
- 03:54.0seenGraphic links a <PricingTable /> Starter card (Active, $9 / month, ✓ Subscribed) to a User panel for “Brooke Millie”; cursor rests on the connector.
- 03:55.7points at“attached to the user itself” = [form “Subscription”] (cursor)
- 03:55.7momentdemo payoff: words, cursor and screen text agree exactly
Fig. A.2 — One minute of video. The transcript, then everything it left out.
Everything the viewer knew, timestamped.
Screensay returns one document per video. The text on screen, exact. Slides, interfaces, code. The words spoken and the tone they were spoken in. What the speaker pointed at. The moments worth remembering, and chapters. Every entry carries a timestamp, so screen and speech line up. Two forms: JSON for your code, and a screenplay, a script-like text that language models read naturally.
JSONfor your code03:14.8
{
"id": "utt_0028",
"t0": 193.4, "t1": 201.7,
"spk": "spk_01",
"text": "Well, it was until I started using
today's sponsor more, Clerk. …",
"delivery": { "read": "promotional",
"contour": "rising" }
}
{
"id": "pt_0011", "t": 194.8,
"phrase": "today's sponsor",
"target": "txt_0301",
"target_text": "clerk",
"how": "screen_match"
}
Screenplayfor models03:14.8
- 03:13.0seenHard cut to a browser filling the frame: Clerk homepage hero; Theo in PiP top-right looks at camera.
- 03:13.4THEO(promotional, rising) Well, it was until I started using today's sponsor more, Clerk. Because not only do they have the best auth platform…
- 03:14.8points at“today's sponsor” = “clerk” (screen match)
Fig. A.3 — One video, two forms. JSON for code, screenplay for models. Same timestamps.
Video has been invisible to language models.
It was never written down. Text, they read. Code, they read. A video is an hour of screen, speech, and gesture that only makes sense together, and its one written record has been the transcript. So models have treated video as speech with a thumbnail. Whatever lived on screen, and whatever was carried by how a thing was said, never reached the answer. When a video becomes a document, that ends. A model can quote the slide. It can find the minute the demo broke. It can say what “this” was. What the viewer knew is on the page, and a page can be searched.
Built for agents.
One call, one video, one document back. For agents there is an MCP server for Claude Code, Claude Desktop, and Cursor. Its tools include watch_video, which returns the whole document, and ask_video, which takes a question about a video and returns an answer. Webhooks send the finished document to your system when it is ready. Your agent does not have to poll.
You
Today
Screensay
- ShaunAsked in Claude Code“What did Theo point at when he said ‘today’s sponsor’?”14:0214:02ShaunAsked in Claude Code“What did Theo point at when he said ‘today’s sponsor’?”
- 14:02Screensaywatch_videoyoutube.com/watch?v=8GRmLR__OGQ14:02Screensaywatch_videoyoutube.com/watch?v=8GRmLR__OGQ
- 14:05ScreensayDocument ready43:15 · 236 beats · 322 utterances · 121 points14:05ScreensayDocument ready43:15 · 236 beats · 322 utterances · 121 points
- 14:05Screensayask_video“…when he said ‘today’s sponsor’?”14:05Screensayask_video“…when he said ‘today’s sponsor’?”
- 14:05ScreensayAnsweredThe Clerk logo on clerk.com, 03:14.8 (screen match).14:05ScreensayAnsweredThe Clerk logo on clerk.com, 03:14.8 (screen match).
- ShaunRead the answerwith its timestamp14:0514:05ShaunRead the answerwith its timestamp
- 14:06ScreensayWebhook deliveredjob.completed → your system14:06ScreensayWebhook deliveredjob.completed → your system
Fig. A.4 — An agent in Claude Code calls ask_video and reads the answer.
Point it at a channel.
Give Screensay a YouTube channel and it keeps watching. Every new upload arrives as a document, with the same screen text, speech, tone, and timestamps as a single request. Nothing to schedule. Nothing to fetch. A research team can follow a conference channel. A support team can index every product walkthrough without anyone pasting a URL. The archive writes itself, one upload at a time.
youtube.com/@t3dotggunder watch webhook set
- Mon 09:12The 30-minute Next.js migration31:04 · 2,180 rowsdocument ready
- Wed 17:40Why our build got 4× slower18:27 · 1,304 rowsdocument ready
- Fri 08:05A proper guide to Fable 543:15 · 2,395 rowsdocument ready
Each document delivered to your webhook the minute it was ready. Nothing pasted, nothing polled.
Fig. A.5 — A channel under watch. Three new uploads, three new documents.
YouTube first. The rest, coming.
Today Screensay takes a YouTube URL. That is the product, and it is what the preview, the API, the MCP server, and channel monitoring all run on. Uploads and private video are coming. Meeting and screen recordings are coming. Live streams are coming. We do not promise dates. Until one ships, it is a plan, not a product, and this page will say which is which.
sources
- YouTubebuilt
Public videos by URL. The preview, the API, the MCP server and channel monitoring all run on it.
- Uploads & private videocoming
- Live streamsplanned
- Meetings & screen recordingscoming
sources
- YouTubebuilt
Public videos by URL. The preview, the API, the MCP server and channel monitoring all run on it.
- Uploads & private videocoming
- Live streamsplanned
- Meetings & screen recordingscoming
Fig. A.6 — Four sources. YouTube is built. The other three are plans.
Try it on a video you know.
Paste a YouTube URL. The first three minutes come back as a document.
Paste a YouTube URLNo account needed. The first three minutes are free.