Filming a fleet that actually works.

Pick a background

You asked for suggestions. Three directions, each composited under the busiest frame in the film — the offline diagram, with its thin connectors and small mono labels — because that is the frame a background can ruin. Contrast cost against the flat ground is measured, not guessed.

Your notes, this round

Scene 6 really happened

I briefed claude-researcher to add three features to the site and coordinate a rotated review. It spawned gift-wrapping, size-compare and reviews into its own group, and told each of them — its words — "four sibling agents will review your work afterwards, so leave the code legible." The shot is five live panes, then the Overview with all five states at a glance. Nothing staged.

Note by note

The demo is real, including the business

One Claude Code session got the brief you see typed on screen. It ran web searches, read live competitor pages, and wrote research.md: Pawtraits by Rocco at $29.40, PopArtYou at $34, Crown & Paw with ~20 SKUs, boutique artists at $500–$4,437, UK fine-art commissions at £1,600–£5,100 — and Dog Artists in London, who won't publish a price at all. Its conclusion: the gap is $180–$450, and "armour only" buys you a reusable asset library plus a provenance angle, because the Met's Arms and Armor collection is CC0 so you can name the actual harness each portrait is based on. It also flagged that "oil painting" is a claim you cannot make if the work is digital-and-printed, and refused to build on the $1.2bn / 28% CAGR market-size figures because they trace to SEO blogs rather than primary research.

A Codex session then read that file and built the storefront. It used the research: the price lands at $340 — inside the gap — and the four armour styles are named after real harnesses with real dates. It tested its own work with Playwright on desktop and mobile before saying it was done.

What the demo actually did

The use case is mine, as you said it could be. I briefed one Claude Code session: “benchmark the fastest way to batch-resize 100 jpegs — use the brief”, with a brief.md attached by drag-and-drop. On its own it then:

1. ran three web searches and checked what was actually installed on the box (Pillow 12.3.0 with libjpeg-turbo; no cv2, no pyvips — so it adapted the plan);
2. generated a 100-image corpus and wrote RESEARCH.md and BENCH_CONTRACT.md for its peers;
3. spawned impl-pillow, impl-opencv and reviewer into its own group through the manager API — and even briefed the Codex session sitting next to it, which I had not asked for;
4. armed a monitor and watched them instead of doing their work.

The research brief you see on screen at 0:42 is that file, rendered by the app's own file viewer. The numbers in the Usage panel at 0:51 are the real cost of making this video.

Shot map

The teardown

26 cuts in 56.7 seconds — an average shot of 2.18 s. But the average is a lie: it opens on 2–4 s shots and crescendos into 0.7–1.3 s shots, then decelerates into a 6 s end card. The reason it never feels frantic is that the ground never changes. Every cut lands back on the same pale sage paper.

Six techniques worth stealing

1. The caption is the prompt. Every claim is shown being asked for, in the product's own input field. Zero lines of marketing copy in 57 seconds.
2. The title lands in a hole in the mix. They pull the bed down for five seconds so a still frame reads as a beat rather than a pause. This means the track has to be chosen before the cut is locked, not after.
3. The product is a physical object. Real macro footage of paper gives the UI weight, scale and a reason to cast a shadow. It also means the film has a texture no competitor's does.
4. One label per shot, four words maximum, in a small UI pill — never a headline.
5. The crescendo is a recap. The "everything at once" shot is earned because each element appeared alone first.
6. Accelerate, then brake hard. 2–4 s → 0.8–1.3 s → a 6 s quiet end card. The deceleration is what makes it feel designed rather than fast.

Our actual design tokens, read off the running app

Typefaces are Geist and Geist Mono; radii 4 / 6 / 8 / 10 px. Whatever direction we pick, the film's own typography should be these, at these radii. Nothing in a release film should be set in a font the product does not use.

1 · Paper Fleet

closest to K3

K3's grammar, transplanted. A desk of off-white and sage paper. One printed index card per session — 47 of them, laid out in a grid, each stamped with its CLI mark and its name in Geist Mono. The only thing that moves on the whole desk is the live status frame, composited onto each card: loops turning, solid accents holding, dim cards sitting still. Real panes arrive as white cards that slide in and cast shadows. Wordmark beat on the bare paper. Crescendo: pull back to all 47 cards at once.

paper · sage · kraft · the product's teal · ink

Why it fits

The product is many small things each holding a state. Index cards are literally that. And the app's light theme already sits on the same warm-neutral axis as K3's sage — the two grounds are 12 % apart in lightness.

Hardest part

There is no camera in this pipeline. K3's charm is filmed paper. We would have to fake the plate — a high-resolution paper scan, or generated stills — and composite in 2D. Convincing at 1080p, thin at 4K, and it will not have the depth-of-field breathing that sells K3's opening.

Cost

Highest. Most of the work is a compositing pass that has nothing to do with the product.

Risk

Reads as homage. Anyone who saw the K3 film will name it in the first five seconds.

2 · Mission Control

recommended

The app's own dark theme, as the entire film. #0b0f13 ground, #2bb3bd accent, Geist Mono captions. One continuous camera (Cursor's grammar) through a single high-resolution capture of the real fleet: open on one status frame filling the screen, its working loop turning; pull back and it becomes an icon on a sidebar row; keep pulling and it becomes 47 rows in mixed states; cut to the Overview and watch a card flip to your turn. Captions are small, letter-spaced, bottom-left, in the product's own mono (Linear's grammar). One graphic device only: a 1 px cyan filament that draws itself from an agent's row to the peers it spawned — Linear's filament, except ours is real, because the operations log has the actual edges.

night · panel · accent · text · muted  — the dark theme, unmodified

Why it fits

It is the money shot the brief names, filmed as itself. And it is genuinely differentiated: the category is white (Cursor) or pure black (Linear). Teal-on-near-black looks like nothing else in it.

Hardest part

Getting 47 sessions into deliberate states at capture time, and having enough visible motion that a wide shot reads as alive rather than as a screenshot. Solved in §04 and §05 — the fleet is choreographed by script, not by luck.

Cost

Lowest of the three. One capture session, camera moves in post, one small graphic layer.

Risk

Dark UI films can read cold. Mitigated by the deceleration and the phone beat, which are the two human moments.

3 · The Handoff

most conceptual

Linear's grammar, but the diagram is our real topology. Black, 2.00:1, no UI for the first 20 seconds: 47 dots, one per real session, each labelled with its actual name in 9 px mono. Edges animate in as the prompts that actually fired between agents this week — sourced from /api/operations — so the graph builds itself as a time-lapse of the week that built the product. Then the camera lands on one dot and it resolves into the real pane, and the last 25 seconds are the product.

black · panel · white filament · one accent, used once

Why it fits

"Agents that talk to each other" becomes a picture nobody else can draw, because nobody else has the log. It is the only direction where the film's central image is a fact about us.

Hardest part

All of the interesting work is a bespoke graph animation — real engineering, not editing. And it is the least product-forward: if the UI does not land hard at 0:20 it reads as an abstract.

Cost

Middle, but front-loaded and the least reusable.

Risk

Highest variance. Could be the best of the three or an art piece that sells nothing.

12 cuts, 5.1 s mean. Act I is one unbroken take — the launcher, both launches and the grouping all happen in a single continuous pull-back, so the film opens the way Cursor's does. The two concept beats at 0:18 and 0:56 are deliberately the same device, which makes the second one read as a callback rather than a new idea.

What we are showing, in order

Cut from rev 01 at your call: needs-input detection, the i panel and trace download, reader search, archive, restart durability, and the 47-session wide shot. They are all still filmable later as separate short clips if you want them for the README rather than the film.

What has to be staged, and how — honestly

Rev 02 is much easier to stage than rev 01, because it never needs a big fleet. Five sessions is the most that is ever on screen at once. But the story now depends on timing instead of scale: the spawn has to land at 0:26 and the review has to start at 0:35.

The app, as captured for this document

Captured from the dev Space at 7e74308 on 19 Aug 2026, 2× device pixel ratio. Six pre-existing sessions — nothing was created.

The capture path

Playwright can record video, but its recorder is VP8 at 25 fps with soft, jittery frame timing — fine for a small inset, not for delivery. The good path is a virtual X display and ffmpeg screen capture, which gives exact frame timing and any frame rate we ask for.

# verified: produced exactly 480 frames in 8.000s at 1920x1080
Xvfb :99 -screen 0 1920x1080x24 -nolisten tcp &
DISPLAY=:99 node drive.js &            # headful Chromium, Playwright-driven
DISPLAY=:99 ffmpeg -f x11grab -draw_mouse 0 -framerate 60 \
  -video_size 1920x1080 -i :99 -c:v libx264 -preset ultrafast -crf 12 plate.mp4

Choreography — one clock for the whole take

Because the dev Space's API answers from this box, a single director script can hold one clock and fire everything on it. That turns "hope the fleet looks busy" into a timeline:

t=0.0   start ffmpeg
t=1.5   POST /api/agents/:id/prompt  x8   # eight rows start turning, on cue
t=6.0   POST /api/agents/:id/stop    x3   # three rows go dim, in shot
t=11.0  page.evaluate(scroll sidebar)     # the pull-back reveal
t=13.2  a pre-timed short prompt lands    # a card flips to "your turn" at 0:13.2
t=20.0  page.click(pane) ; type the real prompt
t=21.6  the parent agent POSTs three peers of its own
...

Camera moves are not filmed. Beats 1–3 are a single static capture of a 1440p plate; the pull-back is a crop expression animated in post. Same for the push into the i panel. This is exactly how Cursor gets 64 seconds with three cuts.

Post

• Camera moves: a Python script emits a per-frame easing table into an ffmpeg sendcmd file driving crop, then scale=1920:1080:flags=lanczos. zoompan is the obvious choice and the wrong one — its per-frame precision is too coarse for a slow pull-back, it visibly steps.
• Graphic layer (title, filament, end card): rendered as an HTML page whose animation is driven by a ?t= parameter rather than the wall clock, screenshotted frame by frame to a transparent PNG sequence. Deterministic 60 fps, zero dropped frames, and it can use the product's real CSS. Composited with overlay.
• Delivery: 1920×1080 H.264 high, CRF 18, +faststart; plus 1080×1350 and 1080×1920 crops with a caption-safe centre; plus a silent 10 s loop for the README.

Seams — the things that will bite

1. --kiosk does not hide Chromium's tab strip and omnibox — verified the hard way; my first test capture has a browser in it. Fix: push the window off-screen with --window-position=0,-80 --window-size=1920,1160, or grab with an offset (-i :99+0,80), or launch via --app=.
2. The dev Space is cpu-basic. Rev 02 only ever needs five sessions, so this is much less of a risk than it was — but five live CLIs on two vCPUs can still stream slowly, and slow streaming reads as a sluggish product. Either bump the dev Space's hardware for the shoot, or use prod with demo mode.
3. Terminal text rendering. xterm.js on a GPU-less box falls back to canvas rendering; glyphs can differ subtly from what you see on your machine. Check one still against a real device before we lock.
4. Font-loading race. The first frames after navigation can show fallback fonts. Warm the page, then roll.
5. No music on this box, and no licence for any. This matters more than it sounds: K3's best trick is a five-second hole in the mix, and you cannot cut to a hole you have not heard. Pick the track before we lock the edit, not after.
6. No pointer. -draw_mouse 0 removes it. Linear shows no cursor, Cursor shows a real one. I would show none — a synthetic drawn cursor always looks synthetic.
7. Every spawned session costs your quota. Nothing has been spawned. I will bring a count and an estimate first.

Rough shape of the work, once a direction is picked

1. Is the K3 film the Kimi one? Still unconfirmed. §01 and §03 both change if it is the Creality printer launch instead.
2. The music. Send a track, or tell me to go find candidates. The title at 0:14 is written to sit in a hole in the mix, and a hole cannot be cut to before it exists — so this one blocks locking the edit, not starting it.
3. Where we shoot, and permission to run the fleet. Dev Space, or prod with demo mode. Either way I will bring you the session count and a cost estimate before anything is spawned.

Two smaller ones I can decide myself unless you care: whether the phone beat is a composited device frame or a real phone filmed on your side, and whether a remote agent is connected before the shoot — without one, the 0:56 beat has to lean on the concept diagram alone, which is weaker.