Product launch video: one command

You finish the project, then stall at the last mile: the launch post needs a video, a screenshot with a caption will not do, and hiring someone is slow and expensive. brag represents the new answer — say let's /brag inside the project directory and the agent reads your code, then renders a roughly twenty-second film through HyperFrames, music and share copy included. The repo was created on June 16, 2026; as of October 11 it holds about 15,079 stars, ships under MIT, and has a hosted sibling that charges $19 per video.[1][2][3]

brag is not an isolated product, though. It is the fastest-moving member of a branch that is only now taking shape. This piece opens up the market for product launch videos: who makes them, what they cost, why a “code-rendered” route appeared in 2026, and the question you should actually ask before choosing a tool.

Key points

The market splits into seven layers: text-to-video models, general AI video platforms, screen-recording polish, device mockups, interactive demos, agencies, and the 2026 newcomer — code-rendered launch videos. The first two make pictures; the last four explain a product.
The code-rendered branch has one shared technical base: the model writes HTML or React, and a headless browser renders it frame by frame. Its selling point is determinism — every word and interface element on screen is rendered, not generated. brag, shipvideo and EndFrame all sit here.[1][6][7][8]
Four billing models get conflated constantly: open source (free, you pay compute), hosted per-video ($19), SaaS subscriptions ($9–149/month), and agencies (one producer’s estimator starts at $6,000). Comparing the raw numbers across them produces the wrong conclusion.[3][10][12][13]

Seven layers of supply

Sort “make the product visible” by deliverable and billing model and brag lands in the narrowest layer. All prices below were checked on 2026-10-11 against the vendors’ own pages.

Layer	Examples	Deliverable	Price anchor
Text-to-video models	Pika and peers	Any generated footage	Credits or subscription
General AI video platforms	HeyGen, VideoGen	Marketing short videos	$19–149/month
Screen-recording polish	Screen Studio	Screen capture with camera moves	$9/month yearly, $29 monthly
Device mockups	Rotato	Design work inside device frames	One-time purchase, no public figure
Interactive demos	Arcade, Supademo	Clickable product tours	Supademo from $50/seat/month
Agencies	Animation studios	Custom film	Estimator starts at $6,000
Code-rendered launch videos	brag, shipvideo, EndFrame	20–40 second launch videos	Free open source / $19 per video / bring your own subscription

The important reading is that the first two layers and the last four solve different problems. A text-to-video model can generate beautiful footage, but it does not know what your product is. HeyGen wraps avatars and templates into a monthly subscription, which suits people who already have a script and assets. The layer that actually answers “how do I explain the thing I just built” is the bottom four rows — and their prices span two orders of magnitude.[7][10][11][12][13][14][15]

The code-rendered branch

The branch that appeared in 2026 shares one technical base: the model does not generate video. The model writes code, and the code renders video.

brag has the agent scan the project directory first and answer a nine-question rubric before planning starts, then write a beat-by-beat storyboard and hand a brief to HeyGen’s open-source HyperFrames (about 60,587 stars as of October 11, 2026, Apache-2.0). shipvideo is more direct: Opus 5.5 writes a single-file HTML film, a headless Chromium renders it frame by frame inside an Amazon Linux microVM, and ffmpeg encodes the MP4 — roughly four minutes and 100k tokens per video.[1][6][7][18]

That determinism is the entire advantage over text-to-video. Launch videos are wall-to-wall text and interface elements, which is exactly where video models fail. Every frame of a code render is a definite output: change a word, re-render, and the result is reproducible pixel for pixel. brag’s own creative laws say at least one scene must show real product UI, and this is the engineering that makes that promise safe.[1]

Smaller experiments line up behind it: EndFrame puts the film on a real timeline, renders actual frames after every edit, and runs a 33-rule quality gate over margins and text contrast, all driven by the Claude, ChatGPT or Grok subscription you already have (free during early access); explainroo (528 stars, MIT) bills itself as a local “your agent makes explainers” tool; motion-video-kit (1,082 stars, MIT) ships a commercial video skill kit with an independent critic loop.[8][16][17]

Watch the calendar, though. shipvideo was created on September 24, 2026, explainroo on September 25, and EndFrame’s Show HN post went up on September 8 — at least four independent implementations in three weeks. The underlying pieces (HyperFrames, Remotion, a headless browser) are all public, which means the technology is not the moat.[9]

How deterministic rendering works

Turning “a deterministic video” into reality requires solving three engineering problems, and shipvideo’s README lays the approach out plainly enough to read as a technical sample of the whole branch.[6]

The first is time. Animations in a browser run on four clocks: requestAnimationFrame, timers, the Date object, and CSS/Web Animations. A renderer cannot wait for real time to pass — that is both slow and unrepeatable. The fix is a virtual clock that takes over all four time sources, plus a __seek(t) call that jumps the page to any moment. The renderer then advances frame by frame: frame 0, 1/30th of a second, 2/30ths, each one an independent, definite page state.[6]

The second is frame output. Each frame is captured at 1920×1080, 30fps, piped as JPEG into ffmpeg, and encoded to H.264 (crf 18, yuv420p, faststart). No GPU is involved: a 30-second film renders in roughly 30 to 40 seconds, close to real time.[6]

The third is self-checking. shipvideo hands the agent a check_scene tool that loads the film under the virtual clock and reports JavaScript errors plus the visible text at several timestamps, so the agent can fix problems before rendering. EndFrame goes heavier: it renders real frames after each edit and checks them against 33 rules covering margins and text contrast. Both mechanisms exist for the same reason — catch the layout problem while writing code, not after rendering a million frames.[6][8]

Those constraints explain why the branch only became common in 2026. Remotion turned video into React components in the early 2020s and HyperFrames turned video into HTML with data-* timing attributes in March 2026; what actually unlocked “the model writes the code” was models finally writing several hundred lines of correct front-end code in one pass — shipvideo’s README cites Deedy’s Opus 5.5 instructional-video demo as its proof. This is not an isolated pattern either: our own review of ten video agent skills found deliverables ranging from MP4s to Lottie files to word-deletion lists, all built on the same public primitives.[6][18]

What an agency’s $6,000 buys

Treating the agency price as merely “expensive” misses the point. It buys a different thing from the tool routes, and understanding the difference is what tells you when to spend it.

One production company’s cost estimator prices on four inputs: goal, video type, quantity, and whether you need a script. 2D animation, motion graphics, whiteboard and mixed media are separate tiers, and within a tier, scriptwriting, voiceover and revision rounds all move the number. That is why quotes can vary fivefold: you are not buying a video, you are buying a chain of human steps.[13]

For a typical engagement the money goes to scripting and storyboarding (which set the film’s information structure), visual style and asset design, animation and compositing, voiceover and music, and one or two revision rounds. For a launch event, a funding announcement, or a homepage hero, those steps cannot be skipped — they provide exactly what the tool routes do not: control and human judgment.[13]

The tool routes save precisely that. brag does not let you rewrite the script, swap a shot or retime a cut; it bets those are unnecessary for a launch post. That is not a capability gap, it is a division of scenes: agencies make the film many people will inspect, tools make the asset you publish the day you finish.[1]

Four billing models

Price is where misreading happens most. The same “one launch video” can be billed four ways, and the numbers do not share a coordinate system:

Model	Example	Per video	Hidden cost
Open-source skill	brag	$0	Your compute; Node 22+, FFmpeg to install
Hosted per-video	letsbrag.app	$19 (listed later price $29)	Same skill underneath
SaaS subscription	Screen Studio, Supademo	$9–50/month/seat	You capture and edit the footage
Agency	Animation studio	Estimator starts at $6,000	Weeks of calendar time; revision process

The agency figure comes from one production company’s public calculator, which returns ranges from $6,000–$15,000 at the entry point up to $25,000–$62,000 at the top. It is one company’s pricing logic rather than an industry average, yet it establishes the scale: a $19 hosted brag is under four-tenths of one percent of that entry price.[3][13]

Be equally clear about the open-source cost. The skill is free; inference and render time go on your own account: Node.js 22+, FFmpeg and the HyperFrames CLI on macOS, with the rendering stack downloading on first run. The hosted refund policy is unusually blunt — no video, full automatic refund; wrong video, one free remake or a refund within 14 days. The terms name the operator as the author himself, with Creem as merchant of record for payments and tax.[3][4][5]

brag’s bet

brag freezes four decisions into rules, and that is both the product and its boundary.

Duration is locked to 15–25 seconds, not one second more without a reason. Content must show real interface at least once, and generic SaaS language like “streamline your workflow” is banned outright. Tone offers seven presets plus freeform direction (“fake Series A launch from 2016”). Structure runs hook, reveal, highlights, punchline. There is no timeline, no re-cutting, no asset import.[1][2]

That bets “one pass is enough.” For a launch post the bet likely holds: nobody watches past 30 seconds on X or Product Hunt, and 15–25 seconds is the shortest length that tells a whole product story. But anyone who needs precise cuts, narration sync or mixed footage is not the user — they want a recorder or an editor, or a template-and-timeline route such as video-shotcraft.[1]

The other asymmetry is distribution. The skill is free and lives inside the user’s own agent; the hosted version charges $19 per video. That structure pushes marginal cost to nearly zero, and the price is that everyone willing to run it themselves never becomes revenue — the README says so outright. The real asset is accumulated creative rules plus reach, not the code: the skills.sh listing, the credibility of 15k stars, and the launch videos a dozen creators on X have posted on their own.[1][19]

Where the Chinese market sits

Chinese demand signals exist, but supply is mismatched.

Bing’s Chinese autosuggest shows clear intent. The query for “product promo video production” suggests “AI-made product promo video,” “software product promo video production” and “how to make a product promo video”; the query for “AI generates product video” suggests “free AI-generated product video” and “AI generates product intro video.” Readers are actively searching for AI-made product videos, so the demand is real.[20]

Supply, though, concentrates on two things: turning e-commerce product photos into video, and avatar-driven narration. Bing’s suggestions for the phrase “product launch video” even lean toward “product launch event video” and “Apple product launch event video” — in Chinese the phrase largely means an event livestream, not a launch asset. The developer-facing position (“read my code, decide the narrative, generate launch-post material”) has no Chinese product and no Chinese search term.[20]

It will probably stay empty for a while: brag’s own SKILL.md is written in English, and so is every sibling implementation’s documentation. That is not a call to fill it. It just means that if you are launching in Chinese today, your realistic choices are English tools or general-purpose platforms.

How to choose

Back to a decision. Ask one question first: which step are you willing to hand to a model?

Narrative and rendering — a skill like brag. You bring the project; it brings the angle and the film. The cost is no editing pass.
The editing timeline — a tool like EndFrame. The model drafts scenes and you revise them on a timeline, if you have a model subscription.
Camera moves and capture — a recorder like Screen Studio. You drive the product while it handles the polish, and you supply the footage.
A production team — budget from $6,000 for full customization and human review.

The second question is where the film will run. A post on X or Product Hunt is fine with a 15–25 second one-pass render; a homepage hero loop or a sales-deck demo needs every frame controlled, and this route is the wrong one.

One last caveat: stars are not adoption. brag’s 15k stars and 1,101 forks are bookmarking signals, and EndFrame’s Show HN post earned four points and no comments. Those numbers measure attention, not how many people posted a launch video made with them.[1][9]

FAQ

Code rendering vs text-to-video

Text-to-video makes a model imagine each frame, and text or interfaces tend to warp. Code rendering has the model write HTML or React that a browser renders deterministically, so output is definite and reproducible. Because launch videos are mostly text and UI, this branch chose code.[1][6]

Does brag only work on code projects?

No. The project mode reads local code; the website mode fetches a live page (including JS-rendered sites), so projects without source work too. Pages behind a login cannot be captured, and the hosted version refunds those orders automatically.[1][5]

Open-source vs hosted output

Per the project, both follow the same creative rules; the self-run version is driven by your own model, while the hosted version runs on the author’s servers. What you save corresponds to your own install, debugging and waiting time.[1][3]

The Product Launch Video Is Becoming One Command