Claude Code skill that makes launch videos, ads and music videos from one line: idea, script, visuals, voice, music, edit and review, then a review room where your notes make the next take
A directed look: when the frame must not look like every other AI explainer
Read this before designing the part of a film that explains a product, and any time someone says a cut "looks like AI" or "like all the others". It came from a launch film whose ad was praised and whose explaining part was not: viewers said it reminded them of every AI video, the same pattern, the same structure. The storyboard was fine. The design was the default.
Why "avoid the AI look" does not work
Asking yourself for "something that doesn't look like AI" swaps one default for another. On that film the first fix replaced cards and pills with a flat drawing of a desk covered in paper, stamps and labels. It was a second default, with no real light, and the owner found it worse than the original.
What works has two halves: name the defaults and ban each one by name, and replace every adjective with a number.
That fixes the styling. It does not fix what is being shown, and on our film that turned out to be the larger half. Read the next three sections before anything about style.
Ways to explain a product, and when each fits
The default look is what comes out when the design is made from memory. These are the ways real launch films explain a product, taken from the films in the next section. Pick the one that fits what this film is for, on your own, and say which and why in the art director's line. It is a way of showing, not a template: the brand, the idea and the film's own material decide what it looks like. Do not show the person other people's films or ask them to choose one.
| Way | What is on screen | Fits when |
|---|---|---|
| The real interface, very close | one control, one row or one typed line fills the frame, on white or black, with a few small words | software with a screen; one feature |
| One object that changes | a list, a board or a page gains and loses parts step by step, one word naming each step | a process, a pipeline, "how it works" |
| The product working | the real canvas filling up, the real output arriving, nothing drawn around it | agents or automation doing work you can see |
| The product's own output as the film | the things made with it, full frame, cut fast | creative tools, libraries, anything whose result is a picture |
| A person, plain shapes beside them | a founder at a desk, squares and lines on black next to them | a new idea that needs trust |
| Life, with the product small | people, hands and places; the interface as a small card or a phone in a hand | consumer products; a feeling more than a feature |
| Mixed media in one colour | footage, type cards, a close look at the interface, cut together in one accent | a strong brand colour and real footage to work with |
| Type alone, with the craft visible | one phrase at a time, thin guide lines and numbers around it | nothing to show yet; a manifesto |
| A few steps, each with its own picture | five lines, five different small scenes, the words kept small | a method of three to five steps |
Two ways can share a film (the product working, then its output full frame). One frame repeated for every item of a list is not on this table.
What the films that look agency-made have in common
From 25 films read frame by frame in October 2026 (pictures only; nobody listened to them).
- They show the real thing, very close. One real control, one real row, one real line of text fills the frame: a button shot at an angle with the far edge out of focus, a file list in plain mono type, the box the user types in. The interface is rebuilt cleanly, then treated like a product on a table. Nothing is explained with invented icons, numbered pills or diagram panels.
- Real people, real footage, real printed things. A founder at a desk, hands on paper, a walk at golden hour, a clip of old home video, a printed specimen book. One real shot in a graphic film changes how the whole film is read.
- A strict palette. Black, white or one paper tone, and one accent colour that shows up a few times in the whole film.
- Small type and a lot of empty space. One typeface, sometimes one italic word as the accent. The words sit small in the middle or the lower third, often typed with a caret. No bold headline in a corner over a panel.
- A different technique in every shot, inside that one system. Footage, then a type card, then a close look at the interface, then a collage, all in the same two colours and one face.
- Nobody lists. No film showed N items in N matching frames. A process is shown as one real object that changes at each step, with one word naming the step (a file list that loses rows, gains folders, grows connecting lines). A team of agents is never drawn; the product is shown working, and three to five moments carry it.
- Short. Most run 20 to 45 seconds in all.
The skeleton is the tell, not the skin
Our explainer was rebuilt twice, as a paper desk and as a lit 3D stage, and both kept the same skeleton: a title with a number, a container, a row of invented things, once for each of twelve jobs. The owner preferred neither to the original ("I don't think we cracked it"). A new skin on the same skeleton is the same film.
So before any restyling, change what is shown:
- For each step ask: what real thing did this step make? A line of the brief, the sketch, the shot, the track, the note, the score. Put that on screen at full size. The film's own real material is its best picture; ours had a praised ad and kept it in a box.
- A step that made nothing you can see shares a shot with its neighbour.
- A list becomes one picture that changes, not one frame per item. Twelve jobs can be twelve changes to a single full-frame shot.
- Keep the storyline and the sound. This is a change of pictures, not of script.
- A made-up brand needs a name nobody else's film is using. Two unrelated films made with agent skills carried the same invented brand name, because a model reaches for the same few names; search a name before you commit to it.
The sections below are for when the film's world really is invented (a story, a song, a product with nothing to show yet). When there is a real thing, it comes first.
1. The ban list
Write it at the top of the page's code and keep it alive. After the first stills, name what the render used anyway and add it.
The look people now call the AI look, as our own explainer wore it:
- an off-white or near-black ground with soft colour clouds
- a card with round corners and a soft shadow, a dark panel on a light ground
- a pill with a number and a name; a row or grid of icon tiles
- a big bold sans headline in a corner, one accent colour on everything
- every element popping or sliding up with an overshoot; everything arriving at once
- karaoke boxes on the words; text that moves while it is being read
- the same frame for every item of a list. Twelve jobs in twelve identical panels is the pattern viewers named.
- also: purple or blue gradients, glow orbs, drifting particles, a spinning product on a turntable, everything centred
2. One committed world
Pick one physical or graphic world that belongs to what the product is, with objects that can carry every fact. A film-making skill became a dark sound stage: a row of twelve director's chairs with the job stencilled on each, a piece of real work beside each one, a lamp over each. Then every fact is an event in that world: storyboard cards drop onto a ledge, a board of slats turns a pencil sketch into the shot, a counter's number wheels roll to the score, paper slips swing down with the notes.
Rules for the world:
- A list is one camera move past stations, each built from different objects. It is not one layout shown N times.
- The name of a thing lives in the world (on the chair's back), and once more in the corner label.
- The real result stays real. The film's own shots are the pictures on the boards and the monitor, moving.
- Flat pictures need something behind them. A card floating in front of a far wall reads as an interface. Stand it on a ledge against a board so it casts a shadow.
3. Numbers, not adjectives
"Premium", "cinematic" and "modern" are where the default lives. Write the decision instead.
| Instead of | Write |
|---|---|
| premium materials | pale wood roughness 0.6; satin metal roughness 0.4, metalness 0.85; paper roughness 0.92 |
| nice lighting | one softbox key per subject, front left and high, about 4 m away, warm; a low cool fill from the whole sky; a cool rim from behind; which lamps are seen in frame |
| cinematic camera | field of view 16 to 24 degrees; 11 degrees down; never parked: a 2 to 5 % creep in every hold; one travel between subjects of 0.7 s |
| smooth motion | the easing by name, springs as a damping ratio, the frame each move lands on |
Motion numbers that hold up:
- Enter with ease-out (
0.16, 1, 0.3, 1for reveals), move with ease-in-out (0.76, 0, 0.24, 1), exit with ease-in (0.7, 0, 0.84, 0) and make exits about half as long. Linear only for rotation and counters. - Springs by damping ratio: 1.0 never bounces, 0.7 to 0.85 is calm, 0.4 to 0.6 is lively, under 0.3 is a cartoon. Position and size may bounce; opacity and colour never.
- A spring as a pure function of time (no simulation, so any frame renders alone):
1 - e^(-z w t) (cos(wd t) + (z w / wd) sin(wd t)), withw = 2 pi f,wd = w sqrt(1 - z^2). - One hero move per shot. Secondary moves go 50 to 70 % as far and start 100 to 200 ms later. Start the next element when the last is a third to half done.
- Text holds 0.3 s per word plus 0.5 s before anything moves it.
- Frames per beat = fps x 60 / BPM. Compute every beat from time and round each one; adding a rounded count drifts.
4. Real light, in code
When the world is physical, flat fills and drop shadows give it away. Build it in a 3D engine on the graphics chip and light it like a set.
Setup. npm i [email protected] in the film's folder, load it with an import map from the local server (serve.mjs), render into a canvas that is not in the page, then draw that into the film's 2D canvas so the words can go on top. renderFrame(t) stays a pure function of t. The exporter already starts Chrome with the graphics chip on.
The method: every frame is the average of N renders (40 is enough for most shots). Each render:
- stands each shadow-casting lamp somewhere else inside its softbox (a 0.8 m square gives a soft edge that hardens at the contact)
- takes one direction of sky for the fill, drawn with more weight near straight up, with its own shadow map: this is what darkens corners and the ground under things
- shifts the camera inside its lens opening and skews the projection so the focus plane stays put: move the camera by
(ox, oy)along its own right and up, thenm[8] = -m[0] * ox / focus,m[9] = -m[5] * oy / focuson the projection matrix - nudges the picture by a part of a pixel (add
-2 * jx / width,-2 * jy / heightto the same two entries) - steps the time inside half a frame, which is the motion blur of a film shutter
Add the renders into a float buffer with weight 1/N, then one last pass: exposure, a curve that is straight to about 0.72 and rolls off above it, a small glow from what is brighter than white, darker corners, grain seeded by the frame number. Use the same sample set every frame, so what is left of the sampling stands still instead of flickering.
What went wrong on the way, so check for it:
- A bright set looked like a basic 3D render. With no bounced light, a white studio goes flat and grey. A dark stage is honest for this method: light falls into black, the pools of light do the composing, and simple shapes read as a set, not as a toy.
- Fill with no shadow washes everything out. Environment light, a hemisphere light and lamps that cast no shadow flattened the first stills. Keep them very low; let the key and the sky do the work.
- A pixel with an endless value turns into black blocks once the glow blurs it. Clamp what you read before adding it up, and drop any value that is not a number.
- Small lights far out of focus show the sample pattern as a cluster of dots. Make them larger than the spacing between samples, or raise N for that shot (128 while a slate filled the frame).
- Rounded boxes and plain boxes will not merge until both are made index-free.
- A mirror in the floor (the engine's reflector, added on top at about 0.2, stronger at a flat angle, blurred a little) did more for the "shot on a camera" feel than any material.
- Every word on a prop is a texture: load the fonts before building the scene.
- A frame at a cut mixed two shots. The renders of one frame are spread over the shutter, so the frame on a cut took half its renders from the next shot. Keep every render of a frame inside the shot the frame belongs to.
- Milky pictures. A small lift added to the blacks "like a print", plus glow, put the darkest pixel at 25 of 255 and made the film's own photos look washed next to the same photos in the old cut. No lift, less glow, a mild S-curve at the end, and measure the darkest pixel of a still.
- Framing. Decide how wide the frame is at the subject (4.6 m here) and place things from that. Nothing the viewer should read may be cut by the frame's edge when the camera settles.
5. Shots: no job gets the same frame as the last
The first cut of the new world scored 5.5 with a fresh reviewer, for reasons that had nothing to do with the light. Each one is now a rule.
- Each job gets its own shot size. A close look at the work (two cards landing on the ledge, the counter's wheels rolling), then the station. "Chair left, panel middle, prop right" three times in a row is the old card template in 3D.
- The real result fills the frame for a beat. The ad's own shot is the best picture in the film: after the board has turned to it, cut in until the board is nearly the whole frame.
- A prop needs its parts. A grey block with two discs is clip art next to a real photo. Give the hero prop what makes it that thing: reels with holes that turn while it runs, bright rings on the lens, a hood, a name plate, a red lamp.
- What the words say must happen. "It goes back" is the camera racing back down the row to the camera station, the shot being made again, and the camera returning to a new score. A change of text on a screen is not "going back".
- Nothing waits empty. A wire with no notes on it yet, a stand with no lamp on it: both read as mistakes. Build the thing that holds the notes as a board that looks finished before they land.
- Vary a repeat. Six cards falling exactly alike read as one animation copied. Change the fall time, the tilt and the settle a little for each, and speed the rhythm up.
- A thing held in front of the lens is held by someone. A slate floating in mid-frame looks wrong. Bring it so close that its lower edge leaves the frame, and keep its bottom row empty so no word is cut.
- A very fast sideways move leaves the renders of one frame as a row of copies. Smear each across the gap to the next one (a sideways blur as wide as the camera slides between two renders).
- Grain from a cheap hash shows a fine mesh that reads as compression. Use an integer hash on the pixel position.
6. Words: the film's own typeface, large enough for a phone
The first try borrowed a label system seen in a reference (a small figure number, a timecode, an italic serif line). The reviewer named where it came from, and none of it could be read on a phone. Use the film's own headline face at about 58 px for the job's number and name, one mono line under it, in the one corner that is empty in every station frame. No words over a close look. The take number sits small in the opposite corner and turns from "TAKE 1" to "TAKE 2" when the cut is sent back.
The second cut scored 6: better, still under the bar. Simple props beside real photos read as toys, and words on top of the picture were still the old film's title. Shown the cut after two rounds, the owner did not prefer it to the original. The lesson was not a third round of props: it was the three sections at the top of this page. An invented world, however well lit, competes with the real material the film already has.
7. Test before rebuilding
Rebuild three moments, not the film: the one that names the product, the strongest pair of steps, and the turn. Put them on the film's own sound and show them next to the old cut. Say what the test does not prove yet: the sound effects were timed to the old picture, the other stations are placeholders, no fresh review has seen it.
Sound, when the picture changes this much
Every hit lands one frame after its picture, never before. Hits sit 3 to 6 dB over the bed in the band a phone can play (about 200 Hz to 6 kHz); a mix can read -14 LUFS and still be quiet there. Check by measuring the decoded file, not by trusting the plan.