Motionmaxxing is a Claude skill that rates and fixes AI videos

AI motion graphics usually fail in the same way, with corner labels, a counter, a left headline next to a card, and a fade between every beat. Motionmaxxing is a new agent skill that calls that failure "mid" and ships six gates to catch it. Tejas Makwana published the project on 6 October. The GitHub REST API shows 193 stars and 22 forks, and the README says 53 pro moments were decoded into about 150 designer rules.

The skill plugs into Claude Code and any other coding agent that reads a SKILL.md folder. A user hands the agent a draft video or a plain brief, and the agent plans the idea, writes a storyboard, builds the film in vanilla HTML and canvas, renders every frame through Chrome and ffmpeg, and then checks the output against the gates.

The rating scale and the gates

Motionmaxxing scores a video on three bands: slop (a slide deck that moves), mid (clean and undecided, with default eases and flat grounds), and maxxed (one idea, measured motion, cuts on motion, and gates that pass with numbers shown). The author built the scale from a slop autopsy of rejected AI films that ranked 25 default patterns, and then wrote the gates to catch those patterns.

Four of the six gates are run by scripts and two are judged by eye on the contact sheet. G0 checks the file decodes and moves. G2 fails any run of more than three flat frames in a row. G3 fails an end hold longer than 1.4 seconds or a final shot that takes more than a quarter of the film. G5 is the page-chrome lint, which fails corner labels, counters, timecodes, progress bars, left headline blocks beside right-side media, and letter-typed headlines. G1 and G4 need a person to look at the sheet.

The measurements behind the rules

Every rule comes with a measurement from the pro corpus. The median move is 10 frames, which the README gives as 0.33 seconds. Landing frames decelerate 58 percent of the time, while leaving frames accelerate 37 percent of the time. The corpus holds 3,392 transform entries against 584 opacity fades, so pros move elements rather than fade them. Hard cuts carry 96 percent of the moments studied. In the 3 to 7 second pro clips, something moves on 99 percent of frames.

The README publishes three before and after comparisons on real brands, Cluely, Spotify and Wispr Flow, under an unofficial demonstration note.

The limits named by the author

The author calls this an early release and says so in the project's known-limits file. The lint cannot see text inside a canvas or an image. Gates G1 and G4 are judged by a person on the contact sheet. ElevenLabs music does not always hit the planned tempo. WebGL renders run at 7 to 11 frames per second on his machine. One rater did the earlier blind A to B rounds that chose the with-skill film 9 times out of 10, so that number is a signal rather than a benchmark.

The skill is Apache-2.0, with bundled GSAP, three.js, opentype.js and Inter keeping their own licences. It lives inside the author's product, FinalMotion. Node 22 and above, Python 3, ffmpeg and Chrome are the requirements.

The pitch for a marketing or product team is that a coding agent now has a measured opinion about an AI video before anyone sees it. When G5 fails, the lint prints the exact pattern the agent must change, and the agent fixes it without the team buying another tool.

Source

motionmaxxing on GitHub