The generic look is a briefing failure, and it is fixable in about fifteen minutes.
Prompting a launch video now takes a minute. Scroll any feed and the results rhyme. Same fonts, same drifting gradients, same synthetic gloss, different logos. Competent and completely anonymous.
The common read is that the models are not good enough yet, and that a better one next quarter will fix it. That read is expensive, because it points the fix at a vendor when the fix belongs in the brief.
When the only instruction is “make a launch video,” a model reaches for the middle of everything it has ever seen. The middle is identical for everyone using it. Vague direction returns the statistical average of the training set every time, and no amount of prompt cleverness moves an unbriefed render off that average.
What did the text workflow already solve?
You fixed this for words years ago and probably never framed it that way. Voice rules. Banned phrases. A style guide pasted into every prompt. Tone of voice enforced at review. Brand as code, applied to language.
Video never received that layer. Every clip starts from nothing and lands on the average, then gets shipped because the deadline is real and the alternative is shipping no video at all.
What does a director actually decide?
A studio never lets a render start undirected. Before a single frame exists, someone has forced the look decisions. What world does this brand live in. What type carries the message. Where the accent color goes, and just as importantly where it never goes. What the thing sounds like.
Those decisions are the whole distance between a film and a demo, and they are exactly what one minute of prompting skips.
video-direction is the studio process we run on our own brand videos, packaged as a free plugin for Claude Code and Claude Cowork and given away. It directs. An open renderer called HyperFrames turns the direction into pixels on your own machine.
/plugin marketplace add revwisely/video-direction-plugin
/plugin install video-direction@video-directionHow does a brand write its look down?
Run /brand-init once. It is a taste-by-comparison interview, roughly fifteen minutes, and almost every question is an A or B pick between two options. No design vocabulary is required at any point, which matters because the person who owns the brand is usually not the person who can name a typeface.
The answers become motion-brand.md, one file holding your world, type, accent rules, style roster, and audio defaults. Every future video reads it. The interview closes by rendering a proof frame in your new brand, so the look is on screen before any real video gets made.
Magnetiz runs its own. The Terminal Design System translated into motion, deep charcoal, Steel Blue carrying meaning rather than decoration, JetBrains Mono and DM Sans, a syntax palette where a color change marks a real change of state. Written down once, and every Magnetiz video inherits it.
What do the six stages catch?
Ask for a video in plain language and the process runs treatment, style, boards, styleframes, build, and QC, with a gate at each one.
The gate that earns its keep is styleframes. Two finished stills get rendered and you pick one before anything animates. Critics check those frames against a reference deck of best-in-class work first, so the version reaching you has already survived a look review. Deciding at a still costs seconds. Deciding after a render costs the render.
Does it hold up across a whole calendar?
One on-brand video is a nice accident. Twelve is a system.
Every finished video appends a row to a style ledger. The next video reads that ledger and varies deliberately, so video two carries the same brand as video one without recycling its composition. Consistency without repetition is the part that breaks when a person does this by memory across a quarter.
What does it take to try it?
A beta user ran the whole path twice in the last week. She had never used Claude Code before. First run, brand guide in, on-brand video out on the first pass. Second run, she handed it a class promo compilation with background music and no text, and the process pulled elements out of an existing clip and timed the insert to the music.
Her assessment was that it looks pretty good and still needs a little work, which is the honest register for a tool at this stage and the reason it is free.
Fifteen minutes for the interview. It installs in Claude Code or Claude Cowork, runs on your machine, and the renders stay local.
See the full process and install it
FAQ
Why do AI generated videos all look the same?
Because an unbriefed render returns the statistical average of its training data, and that average is identical for every person prompting it. The model was never given the look decisions that distinguish one brand from another, so it defaults to the middle of everything it has seen. The generic result is a briefing failure rather than a model limitation.
How do you make AI video match your brand?
Write the brand down in a form the process reads before every render. That means a single file holding world, typography, accent rules, motion style, and audio defaults, plus a process that consults it at the style stage rather than hoping a prompt carries it. Brand consistency in video comes from a persistent brand file, the same way it came from a style guide for copy.
What is a motion brand system?
A motion brand system is the video equivalent of a written style guide. It records what a brand looks like in motion, which world it lives in, what type it uses, where the accent color is allowed to appear, and what it sounds like. It exists so every video inherits the same decisions instead of re-deciding them under deadline.
Do you need design experience to use it?
No. The brand interview is built as A or B comparisons, so it asks which of two options feels right rather than asking for a typeface or a hex value. A beta user with no prior Claude Code experience completed it and rendered an on-brand video unaided.
What does the styleframe gate do?
It puts a human decision in front of the expensive step. Two finished still frames get rendered and reviewed against a reference deck, and a person picks one before any animation begins. Catching a look problem at a still costs seconds, while catching it after the render costs the render and the review cycle around it.
How does it keep a series of videos from repeating?
Every completed video writes a row to a style ledger, and the next video reads that ledger before choosing its approach. It varies composition and pacing on purpose while holding the brand constants fixed, which is what keeps a quarter of videos recognizable without making them interchangeable.
Does it work in Claude Cowork?
Yes. The same two install commands work in Claude Cowork, and the fifteen-minute brand interview runs there end to end, including the proof frame it renders in your new brand at the finish. Cowork is the easier starting point for anyone who does not spend the day in a terminal.
What does it cost?
The plugin is free and open. It installs in both Claude Code and Claude Cowork, runs inside your own Claude subscription, and renders locally through an open renderer, so the videos and the brand file stay on your machine.