✱ Recent update — hyperframes team @ singapore

Make our builders' video stand out with HyperFrames, consistently.

Feed the prompt. · Define your style. · HyperFrames handles the rest — one-shot, with your agent.

our builders

build with ease · learn with the community

stand out

their style, with taste

consistently

not cherry-picked — consistent, high-quality results

Tweet: official Claude connector announcement

Official Claude connector · 738.5K views

Now live on Grok · 702.6K views

✱ best practice · skills

A best-practice skills set, with routing.

Our builders know how to build — they just don't know how to sell or make a good video. So HyperFrames is easy to build & launch: they create videos by using their agent, not by hand. The goal: feed the prompt, define the style, and HyperFrames handles the rest.

any input → /hyperframes main skill · router → one of these workflows

✱ beat-driven-promo · bgm → video

Drop a track — the promo cuts to its beat.

We reverse-engineered our BGM golden samples into a composable system — 9 group templates over 37 motion primitives. The agent reads the track's beat grid and lands motion on every onset.

♪ bgm.mp3 + › make a promo video about hyperframes bgm_test_2 · 129 BPM · 77 onsets · 23.0s · reversed into 9 templates · 37 primitives

With templates + primitives · hyperframes/

beat-timeline129 BPM · 23.0s

sections

VOID e=0.13

MEDIUM e=0.51

HIGH e=0.80

beat grid

events

rolls

369 →SURGE517 →SURGE323 →DROP

silence

✱ beat-driven-promo · another track

Same system — HeyGen Brand Hit.

A short, punchy cut driven by the same reverse-engineered beat grid — a held void breath snaps into a roll cascade, builds through a dense bridge, then a peak surge into a hard-stop break.

♪ heygen-brand.mp3 + › make a brand hit

heygen-brand · 92.3 BPM · 59 onsets · 12.4s · same 9 templates · 37 primitives

beat-timeline · heygen-brand 0.0s ──────────────── 12.4s

sections

VOID

ROLL

BRIDGE

SURGE

beat grid

events

rolls

386330 →DROP275 →SURGE358

silence

VOID HIGH e=0.82

With templates + primitives · heygen-brand/

✱ website-to-video · any site

Point at a website — get its video.

A website-capture skill reads the live site and pulls its real design language, so the video looks like the brand — not a generic template.

› spotify.com and make a video for this site

Built from spotify.com — its type, color and components, captured live.

website-capture skill

✱Reads the live page and extracts what makes it look like itself.
✱The video inherits the site's own brand system, not a default look.

fonts assets · logos & imagery components & layout

✱ embedded-captions · your video

Bring your own video. Caption it with style.

Hand us an existing video and a caption style — captions are styled, timed, and embedded into the footage, sitting behind the subject instead of covering it.

🎬 my-video.mp4 + › add captions with <style>

AI captions embedded into real footage — demo by Jenson.

what's inside

✱Video matting model — separates subject from background, so captions sit behind people, never blocking faces.
✱Caption-layer compiler — turns a style spec into timed, styled captions burned cleanly into every frame.

✱ our agent · first & only video MCP in the ecosystem

Build with ease — low cost, high efficiency.

We're the first & only MCP in the ecosystem for making videos — including Grok and Claude. While we deliver high-quality results with our best-practice skills, we also offer a low-cost, high-efficiency self-host agent that produces solid results for our users.

Official @claudeai connector · 738.5K views

Now live on @Grok · 702.6K views

✱ claude_mcp_agent · live on Claude & Grok

Any answer → a polished explainer video.

prompt

Grok › make a video about spacex ipo

→

plan

Story + visual design

→

produce

tts bgm captions frame.md·design scene-builder asset-scout

→

render

asset_scout — on-demand real-world asset engine

search → rerank → preprocess (background removal)

→ an atomic figure / logo, or a suitable background image — real assets, never placeholders.

generated < 5 min ~$0.5 USD / min ~1k projects · first 2 days

Real user reaction — @talakoubali on this SpaceX video, made by Grok → HyperFrames.

✱ consistently · eval-driven

How a new skill goes live.

Every new skill or feature comes online through an eval-driven process, run end-to-end by our agents — four steps:

01 Understand The need from our builders — including ourselves.

02 Reverse Take the most representative video and reverse-decompose it into skills + input.

03 Generalize Build a benchmark; solve across more inputs until it generalizes.

04 Eval Pairwise / pointwise comparison or judge — until the skill is verified.

Example next: circle highlight.

✱ eval-driven · example: circle highlight

Worked example: the circle highlight.

Circle / ring highlight is one of the most common asset animations. The hard part is location — cloud models (Claude Opus included) are unreliable at exact pixel coordinates, so the ring drifts off the subject. Here's how we ground it.

The fix — a numbered-grid locate loop · RSVP grid-index, on Opus

The model picks grid strips (columns & rows), never raw coordinates — then crops and grids again to lock on.

Generalizes — Opus across a 7-case test set · native box (red) drifts → grid box (green) locks on · 21.3% → 3.7% avg

before · eyeballed

ring off-target · 14.2% center error

after · grid-located

locked on the logo · 2.4% center error

✱ consistently · beyond skills

We track the models, too.

Beyond our own skills, we track the fundamental models — and monitor how they change. The focus right now: code-to-video — how well a model writes the code that makes a video.

We've collected a large set of real online-video cases — a strong differentiator. Some models already clearly outperform the rest; it's a continuous watch.

code-to-video bench — one prompt → many models each writing the code, scored against a **golden** reference.

win-rate **leaderboard** — some models clearly lead.

✱ hyperframes team @ singapore

Thank you.

Feed the prompt · define your style · HyperFrames handles the rest.

Wenbo · Miao · Jenson — HeyGen all-hands 2026-06-19