The Lab
Short write-ups on the repos, agent skills, pipelines and prompt tricks we’re finding right now. Every post credits and links its source. Anything marked Spotted, not yet tested is research only; it moves to tested once we’ve rendered it ourselves. Posts marked Based on the creator’s video were written from the full video transcript, with misheard names corrected, and haven’t been tested by us either.
- Spotted, not yet tested#motion graphics#agent skills#open sourceHyperFrames: HeyGen's bet that agents should write video as HTML
An Apache-2.0 renderer that turns plain HTML, CSS and GSAP into frame-exact MP4, with a skill router that picks the workflow for you.
Source: HeyGen
- Spotted, not yet tested#motion graphics#self-review#sound designclaude-motion teaches an agent taste, then makes it check its own frames
Motion rules with actual numbers, a contact-sheet review loop, and a sound engine written in plain Python.
Source: @whaleyxbt
- Spotted, not yet tested#Higgsfield#MCP#UGCHiggsfield inside Claude: MCP, Soul ID and Katana
One connector URL turns Claude into a front end for Higgsfield's models, characters and audio. Katana adds reference-driven edits.
Source: Higgsfield
- Spotted, not yet tested#UGC#image-to-video#techniqueAI UGC: spend half your time on the starting frame
The cheapest quality upgrade in AI UGC happens before you press generate.
Source: @maxxmalist
- Based on the creator’s video · not yet tested by us#motion graphics#Higgsfield#Seedance#launch videoA two-layer launch video: code motion over a Seedance plate
Recreate a reference launch video's style for your own product, then layer that motion over cinematic footage built in a strict four-step order.
Source: Jason Lee
- Based on the creator’s video · not yet tested by us#roundup#open source#agent skillsTen GitHub repos from one creator's roundup, checked where we could
An open-source 'Adobe suite', a second Photoshop rebuild, an agent browser, a component MCP and more. We've linked only what we could confirm.
Source: Jack (Jack Roberts' channel, by the self-introduction)
- Spotted, not yet tested#editing#agent skills#open sourcevideo-use: drop raw footage in a folder, get final.mp4 back
Browser Use's open-source editing skill reads your video as a transcript instead of watching it, then cuts, grades, captions and checks its own render.
Source: Browser Use
- Spotted, not yet tested#agents#pipeline#open sourceOpenMontage: an open-source agentic video production system
One of the most-starred agent video repos right now, describing itself as 12 production pipelines with 100+ tools.
Source: calesthio
- Spotted, not yet tested#motion graphics#agent skills#ReactRemotion + remotion-dev/skills: React video your agent can write
The most established code-to-video framework now ships official agent skills. Check the license before you build a business on it.
Source: Remotion (Jonny Burger & team)
- Spotted, not yet tested#agents#UGC#guardrailshypit: 'clone any viral video' with agents, and why we're cautious
A fast-rising repo that rebuilds a viral video's whole workflow, including face swaps. Useful to study, risky to use carelessly.
Source: hypit-ai
- Spotted, not yet tested#agent skills#product video#Remotionvideo-shotcraft: 152 cinematic shots as a Claude Code skill
A skill for Claude Code and Codex that builds cinematic product videos in Remotion from a library of 152 shot types.
Source: Vincentwei1021
- Spotted, not yet tested#Seedance#scripting#workflow#reconstructionSherpa to Seedance 2.5: how a 4.1M-view AI film was likely made
The post credits only 'Script: Sherpa by Pocket FM, Video: Seedance 2.5'. This is our reconstruction of the probable pipeline, not the creator's confirmed process.
Source: @maxxmalist
- Spotted, not yet tested#roundup#open source#rankingsThe top GitHub video repos right now, ranked by stars (Oct 2026)
Twenty repos for AI video, agent video skills and ComfyUI workflows, pulled from the GitHub search API on Oct 14 2026.
Source: Each repo's maintainers (linked)
- Based on the creator’s video · not yet tested by us#animation#explainer#pipeline#promptingFree video from code: the canvas + Playwright + FFmpeg pipeline
No Remotion and no generator credits: Claude writes a function of time, a headless browser draws each frame, FFmpeg stitches them, and the agent checks its own frames.
Source: Chase AI
- Spotted, not yet tested#animation#3D#creatorsKevin Ngo's JavaScript-drawn films, and the 2D-to-Blender handoff
Short story films where the model draws every frame in code, and a neat trick for turning a 2D cut into a 3D one.
Source: Kevin Ngo
- Spotted, not yet tested#agent skills#motion graphics#roundupThe 15-repo motion design stack, sorted by job
A viral list of motion-design repos, which we checked one by one and grouped by what you're actually trying to ship.
Source: @Mnilax
- Spotted, not yet tested#prompting#Three.js#motion graphicsWhat 418 shared prompts teach about one-shot code video
Skillry collects creators' prompts and remakes each video side by side. The prompts that work share a structure.
Source: Skillry (curating creators on X)
- Based on the creator’s video · not yet tested by us#Seedance#workflow#guardrails#costSeedance draft mode, plus a review gate worth copying
Two cost-control habits from Sirio's video: iterate in 480p draft mode and finalise once in 1080p, and refuse to spend until a human has approved the inputs.
Source: Sirio
- Based on the creator’s video · not yet tested by us#motion graphics#brand#workflow#webPull a brand from a URL, then animate it: Jack Roberts' Claude Design levels
Scrape the real logo, colours and fonts with Firecrawl first, so the agent animates the brand instead of inventing one.
Source: Jack Roberts
- Spotted, not yet tested#ComfyUI#Wan#consistencyWan 3.0 in ComfyUI: reference tags for images, video and audio
Official templates for text-to-video, image-to-video and reference-to-video, with clips up to 30 seconds and synced audio.
Source: Comfy Org
- Spotted, not yet tested#music#audio sync#open sourceAn open-source audio stack for video: music, foley, lip-sync, timing
Five repos that cover most of the sound side of an AI video pipeline without a subscription.
Source: ACE Studio & StepFun
- Based on the creator’s video · not yet tested by us#consistency#Google Flow#image-to-video#promptingGoogle Flow and Nano Banana 2.1: a five-prompt consistency routine
Create the subject, move it twice, edit one detail, lock a product, then animate the best still. Plus the limits Google itself lists.
Source: Julian Goldie SEO
- Based on the creator’s video · not yet tested by us#Higgsfield#motion graphics#business#MCPSix motion styles clients pay for, according to Higgsfield
A vendor's map of sellable formats, with the workflow tips worth keeping: storyboard first, HTML design sheets, and a credit tally that includes failed takes.
Source: Higgsfield AI (presenter: Adil)