Product Demo

Product demo videos

One pipeline, three layers: the story (what the clip shows), the motion (how the interface behaves inside the frame), and the title (the two seconds that open it). Reference implementation: src/compositions/Demo.tsx + src/motion.ts + src/title-card.tsx + scripts/render.sh.

New clips copy that pipeline. Do not invent a second one.

1. Settle the brief. Read what you can, ask what you can't.

Do not start a composition until the four things below are settled. But settled is not the same as asked — and asking for what you could have found yourself is the fastest way to make a person regret starting.

Read before you ask

If you can reach the product — its repo, its running app, its live site — read it and propose the whole plan. Take the running order from the marketing site and the screens from the app; neither answers the other's question, and question 2 below has the measurement showing why. Then put the plan up and ask for one confirmation.

Say that describing the flow beats letting you guess

Tell them this in your first message, in one line. Reading the product gives you a plausible flow. It cannot give you the right one, because the code does not record which path converts, which step people get stuck on, which feature shipped last week and needs the attention, or which screen the founder is quietly embarrassed by. Only they know that, and most people do not volunteer it because they assume the agent has it covered.

If you already know the flow you want — the steps, in order — say it and the clip will be much closer to what you have in mind. If not, I'll read your product and propose one.

It is an offer, not a gate. If they do not answer it, read and propose as normal; do not ask again. The point is that they were told the option existed before you spent their time on a plan built from inference.

The same applies to anything else they already know: a line of copy that has to appear, a screen that must not, a feature that is being deprecated. Invite it once, up front, then get on with it.

Ask outright only for what is not in the artifact:

• what the clip is for — a launch, a section nobody scrolls to, an onboarding step. Intent is not in the code.
• which file the ending shows, but only when the ending is media the product generated and the disk offers several with nothing to choose between them. An ending that is just the finished screen needs no asset and no question — see question 3.

Everything else — what is on screen, which flow, which features, how long — you can draft, and drafting it is your job. Propose, mark what you inferred, and be specific about what you would change on a "no". A person correcting a concrete plan gives you better answers in one line than an interview extracts in four.

Never generate an asset to stand in for a payoff, and never substitute a lookalike. Proposing a real file that already exists is not inventing; making a new one is. If the disk is empty, that is the moment to ask.

If they answer in fragments ("dashboard, they filter it, end on the published report"), that is enough — translate it into keyframes yourself. Never hand them a form to fill in, and never ask them to mark up the composition.

What has to be settled, however you get there

Question zero: the whole product, or one feature?

Ask it in plain words, not as a taxonomy:

A demo of the whole product, or a scene about one feature?

A product demo strings three to five features into movements on a through-line. The viewer leaves knowing what the product is.

A feature clip walks one flow start to finish. The viewer leaves knowing how to do one thing.

They are not the same clip at different lengths. A product demo cut as a feature clip shows one capability and hides the rest; a feature clip stretched into a tour skims four things and teaches none.

Then infer where it lives — and say so, don't ask

The answer above predicts the destination almost every time. Take the default, state it in one line with the plan, and let them correct it with a word:

Both crossings are real and both happen — a 42s tour living in a page row, a feature walkthrough posted as a standalone short. That is why you say which one you assumed. What you must not do is assume silently: a launch film built as a page loop comes out short, silent and ending exactly where it started, and nobody names the cause — they just say it feels slight.

"Say it, don't ask it" is about the shape of the exchange, not about hiding the choice. Offering it as a pre-selected option with the consequence spelled out is saying it — see "Put it up as one block" below. What is banned is the open question with no default, which makes them do your work.

1. What is being made on screen? One sentence. Not the product feature — the thing the viewer watches get built. "A revenue report", "a UGC ad for a supplement brand", "a deploy going out". If this is vague the clip becomes a montage of output, which proves the product can generate something but never shows what the visitor would actually do.

2. What does it walk, in order?

Feature demo: the real steps of the one flow. Paste URL → binder. Pick face → lock sheet. Query → filter → publish.

Product tour: which features become movements, in what order, and what carries the viewer between them — a piece of state that survives the cut, a line of copy, one character who reappears. A tour with no through-line is a slideshow of screenshots.

Either way the copy comes from the live product, not invented. Getting this wrong is the only defect that cannot be fixed in the edit.

Read both the site and the app. They answer different questions.

The marketing site tells you which features matter and in what order — the running order of its sections is the company's own ranking of what it sells, and its headline is the promise the clip has to keep.

The app tells you what those features actually look like: the real screens, the real steps, the real copy on the buttons.

Neither alone is enough, and the failure modes are opposite. Build a tour from the site and you get a montage of claims with no screens under them. Build it from the nav and you tour whatever happens to be in the sidebar — Settings, Library, Analytics — while the three things the front page leads with go unmentioned. This has been measured: a tour drafted from a sidebar alone opened on the product's third priority, skipped two of its four headline features, and included a screen the front page does not mention.

So: take the running order from the site, take the screens from the app, and say which came from where when you propose it.

3. How does it end? Usually you already know, and should not ask.

The default: it ends on the finished thing. The flow completes and the last state stays on screen long enough to be read — the report published, the sheet locked, the short exported. That is drawn by the composition like every other frame; there is no separate asset and nothing to request. Hold it 2–3 seconds. A clip that cuts the instant the last button is pressed feels like it was interrupted, and the viewer never sees the thing they were promised.

Then: a page clip folds back to the dark ground it opened on (hard rule 5); a standalone one may end on a card, a logo, a CTA.

The exception, and the only time to ask: when the payoff is media the product generates — a rendered video, a generated image, a character. Then it has to be a real file that already exists. Never generate a stand-in and never substitute a lookalike, because a fabricated payoff misrepresents the one thing a viewer is actually judging: output quality.

That case is also where these clips lie. A clip arguing "same face in every scene" that ends on a different person destroys its own claim in the last four seconds, and no edit repairs it. If the payoff is generated media, check that the thing in the ending is the thing from the flow before you shoot.

Put it up as one block, not as an interview

Everything above is what has to be settled. This is how to put it, and the form matters as much as the content: three prose questions in a row read as an interrogation, and people answer interrogations badly.

If the harness has a structured question tool — Claude Code's AskUserQuestion — use it. One block, three questions, each option carrying a one-line consequence, and your own inference marked (Recommended) and listed first. Everywhere else in this section says state your assumption and let them correct it with a word; a pre-selected recommendation is that, done properly. It is not an interview — it is your plan, pre-filled, one click from being corrected.

Without such a tool, ask the same three in one message, marking which answer you would take if they say nothing.

The three, and nothing else:

Mark the recommendation on Where from the answer you expect on Scope — a tour is nearly always standalone, a feature clip nearly always a page loop. Both crossings are real, which is exactly why it is offered rather than assumed silently.

Flow is the §1 invitation in a form people actually answer. Recommend "read my product" — that is the honest default — but the option to describe it has to be visible, because the code cannot tell you which path converts or which screen the founder is embarrassed by.

Never put length or pace in that block. Both get answered "short" and "fast" on reflex, and a demo that rushes the steps it exists to show is the result. Derive the number, show the arithmetic, name the alternatives — the section below is how.

A fourth question only when the ending is media the product generates and the disk offers several files with nothing to choose between them. Options are the real filenames. Never a generated stand-in.

Do not add a question for the title, the pace, the palette or the delight budget. Those are yours to decide and to say in passing.

The three with defaults — state your choice, don't ask

Decide these yourself and say what you picked in one line. Only ask if the user has already shown they care.

• Title text. Defaults to the headline of the section the clip sits in, with a kicker drawn from its body copy. 2.2s.

• Length. Derive it from the flow, never default to it. Budget one step at the pace below, add 66 for the title and ~30 for the outro, and set <CLIP>_LEN to the total.

  pace	simple step	movement with its own beats
  punchy	70	170
  standard (default)	100	240
  cinematic	140	330

  A simple step is one action and its answer — a click, a panel arriving. A movement has internal beats: a list assembling row by row, a belt rotating through seven formats.

  Show the arithmetic, don't just announce a number. A number with no derivation reads as arbitrary and the user has nothing to push against:

  Four steps and a reveal, standard pace: 66 + 4×100 + 240 + 30 = 736 frames, 24.5s. Say "punchy" for ~18s or "cinematic" for ~32s.

  That line is the whole interaction. Do not ask which pace they want — pace is taste with no right answer, and offered as a question it gets answered "fast", which is how a demo ends up rushing the steps it exists to show. State the default, name the alternatives, move on.

  In a page section: a feature demo lands at 14–25s, a product tour at 35–60s, and clips sharing a row should share a rhythm — match a sibling if there is one. Standalone: the ceiling lifts to 90s, because nobody is watching it for the fifth time.

  If the total falls outside the band for the kind you were told, say so and check — usually it means a step was missed or a movement is really two.

• Where the delight budget goes. One hero beat per phase. Name which element the phase is about and spend it there.

Then show the shot list, and wait

Before you write a line of a composition, put the beat sheet back in the chat. Not code — the plan. One line per phase: the frame it starts on, what happens, and which single element carries that phase.

t0    title     "Acme Studio" · "Ask a question, publish the answer"
0     query     types "Revenue by channel, last quarter", hits Run
96    results   five channel rows stagger in, the counter lands on 1,940
184   filter    "Paid only" lights, three rows dim, the total recounts
274   publish   press, the confirmation springs into its slot
380   out       folds back to the dark ground

Four steps and a hold: 66 for the title, four steps just under the standard 100, and 34 to fold out — 480 frames, 16s. Ends on the published report. Say "punchy" for ~12s.

That list is src/compositions/Demo.tsx, frame for frame. Open it next to this and every number lines up.

Then stop and let them read it. This is the only cheap moment left: a wrong flow costs one message here and a rebuilt composition later, and §1 question 2 is the defect that cannot be fixed in the edit. A correction at this point is someone moving a line in a list.

Keep it to the phases. Do not list every entrance and colour ramp — a shot list nobody finishes reading is not a checkpoint, and the point is that they answer.

Then read the screen you are about to rebuild

The shot list says what happens. This decides what it looks like, and it is the step that determines whether someone who uses the product every day recognises their own software. Skip it and you get this kit's template wearing the product's name in the title bar: plausible, generic, and not theirs.

Nothing that appears on screen may be invented. Not a row, not a card, not a label, not a count, not a colour. If you cannot find where something on screen comes from, you have not finished reading — that is not a licence to write a convincing version.

First, adopt the palette. It is one command and it is not optional.

npm run adopt                          # finds the stylesheet
npm run adopt ../src/app/globals.css   # or point at it
npm run adopt ../src/app/globals.css --curves

This rewrites src/theme.css from the product's own tokens and, with --curves, copies its cubic-bezier easings into src/motion.ts. Run it before you draw anything, and read the report it prints — a ! line means that token fell back to this kit's placeholder, and a placeholder purple next to a real brand colour looks less like the product than plain grey would. If something falls back, find the product's name for it and add it to ALIASES in scripts/adopt-theme.mjs rather than hardcoding a hex in a composition.

Two failures to expect. A product whose theme lives in JS rather than CSS — tailwind.config.ts, a tokens.ts — gives the script nothing; read the values and write src/theme.css yourself, keeping the kit's token names. And a stylesheet that defines only a light palette: the script says no dark block found, and you then owe the user a sentence about which stage the clip runs on, because a light app on the dark ground the title card and the outro sit on is a luminance cliff at both ends.

Do not skip it because the placeholder palette looks fine. It does look fine — it was designed to. That is exactly why demos ship in it.

Then four things to find, in this order.

1. The screen's own source, and what it actually renders. Start at the route and follow the component tree to the end. A route is often a thin wrapper — /app/studio/marketing can turn out to be another workspace opened with a prop, and the thing you are drawing lives two files away.

2. The data behind every list on screen. Rows, cards, chips, steps, tabs and empty states almost never live in the page. They sit in a constants or meta module, and that module holds the real names, the real subtitles and the real order. Take a string you can see and grep it back to where it is defined:

grep -rn "Brand guidelines" src/ | head

Then use the whole list, in its order. Drawing five of eight items is the same defect as inventing five — the viewer counts.

3. The theme this surface runs in. Light or dark is a property of the screen, not of the product: a marketing page and an app shell routinely disagree, and one product can hold both. A product also usually has more than one accent — taking --primary because it is first in the stylesheet is exactly how a purple surface comes out green. Read the component and see which token it reaches for.

4. The arrangement, out of the markup itself. Column count, group headings, the anatomy of a single card — index, title, subtitle, state — and the order they sit in. This is the part that makes it their screen rather than a screen, and it is the part an agent skips, because reading a component tree for layout is slower than inventing one that looks reasonable.

It is all there. A grid class carries the column count. Padding, gap and the type scale carry the density, which is most of why a rebuild looks like a different product even when every label is right. Nesting carries the grouping. Read the classes on the elements you are drawing and resolve them — a grid-cols-2 gap-4 p-6 with text-sm text-muted-foreground under a text-lg font-semibold is a card you can rebuild exactly, and guessing at it is how eight documents in two columns became five rows in one.

The case, and it is this repo's own. An agent shooting Athana's Marketing Studio pulled the button labels out of the code — Render, Rendering scene…, Describe the shot you want first — and stopped there. It then drew a five-row "Brand binder" it had made up, on a dark ground, in the product's green. The real binder is eight documents with their own subtitles, sitting in src/lib/brand-knowledge-meta.ts; the surface is light; and it is purple, because it belongs to Mira and Mira has her own token. All three facts were in the repo the agent was already reading. The clip rendered clean, passed the gate, and was worthless — the same product's in-house pipeline, given the same codebase, had reproduced all of it.

Do not ask for a screenshot. Everything a screenshot would tell you is in the markup, and asking for one is how an agent gives itself permission to skip step 4. The arrangement is the JSX; the density is the padding, the gap and the type scale on those elements; the ground and the accent are the classes and the tokens they resolve to. Read them. It is more work than glancing at a picture and it is the work.

A screenshot is worth asking for in exactly one case: the screen only exists once real data is in it and the shape of that data is not in the repo — a chart of the user's own numbers, a feed of their own posts. Ask for that one screen, name why, and read the rest.

A reshoot skips all of this

Fixing flicker, changing a zoom, recutting an outro on a clip whose plot is already locked — go straight to work. A new clip or a new story does not skip it.

2. What the viewer sees

Two axes, both settled in question zero (§1), and they multiply:

Shot at 1600×1000, delivered 1536×960 (8:5) by default — a wide tile for a page. A standalone clip is usually better at 16:9; change the two numbers in Root.tsx and the scale filter in render.sh together.

Whichever cell you are in, the viewer watches the thing get made. Not a teaser, and not a montage of output.

Length follows the content, not a template. A clip cut to a fixed length either rushes its steps or pads them, and both read immediately.

A page clip has to work with the sound off — it autoplays, and autoplay with audio is blocked outright. A standalone clip plays because someone pressed play, so sound is a real option there; either way it is a second pass over the finished file (§7), never a dependency of the picture.

3. Motion — the interface must not teleport

A clip reads as a screen recording when states snap between frames. Full vocabulary and rationale: references/motion.md. The rules:

1. A boolean in a style is a bug. on ? "var(--brand)" : "var(--border)" snaps. Anchor a ramp at the frame the state changes and pass it through mixToken.

2. Never scale(0). Entrances start at scale(0.9–0.97) and travel ≤16px. At delivery scale a 40px entrance reads as a slide deck.

3. Counters count. A number that jumps is the loudest teleport there is, because the eye is already reading it.

4. Typed text types. Static copy in a field the flow says is being filled reads as a screenshot.

5. Nothing enters as a group. 2-frame stagger minimum.

6. Springs for landing, ramps for travel. damping 14 or above.

7. Weight and layout stay stepped. fontWeight and fontSize cannot tween without reflowing the row every frame.

8. One hero beat per phase. If everything is animated, nothing reads as animated.

9. Lay out to the bottom of the canvas. Panels sized by habit rather than by the frame leave a quarter of the picture empty, and dead space at the bottom of a shot reads as a mistake rather than as restraint. Work out where content has to end — canvas height minus a margin — and size the panels to reach it. At 1600×900 with a 108px header that is about 852.

10. The pointer must land on what it presses, and stop there. Two separate failures, and both have shipped.

    The coordinate. Take every target from a rendered frame, never from reading the CSS. Estimating a chip's centre off padding and font size put the cursor 94px away, on the counter above it — and nothing downstream noticed, because the press fires, the state changes and the frame is clean. §6 is how you catch it.

    The hold. trackPos eases from one key straight into the next, so a target with a single key starts leaving the frame it arrives on. The click beat lands ~20 frames later, by which time the pointer has drifted toward its next target — 21px onto the bottom edge of the chip, which reads as the cursor missing what it just pressed. Give every target two keys, the second at CLICK + 8, outlasting press(). A real pointer stops, presses, and only then goes. The drift scales with the distance to the next target, so the worst miss is always the beat before the longest move — and a target that happens to be followed by a key at the same coordinates hides the bug completely, which is why it went unnoticed.

Everything comes from src/motion.ts. If a curve or duration is missing, add it there so every clip inherits it — never hand-roll one in a composition.

4. Titles — the cold open

Full rationale: references/titles.md. The rules:

1. Words, not letters. A 15-glyph title per-letter is still arriving at frame 30 of a 66-frame card, and boxing each glyph destroys kerning. Per-letter only for a 4–5 glyph wordmark.
2. Mask, never fade. The word travels out from behind a hard edge — overflow: hidden on the line box, word at translateY(110%). Do not also animate its opacity: fading makes the word visible above the edge it is supposed to emerge from, and the mask stops reading.
3. Never scale type. See hard rule 3 below; it binds hardest here.
4. The title leaves before the app arrives, cross-dissolved on the shared dark ground.
5. Kicker follows, never accompanies.
6. Set the font explicitly — src/fonts.ts. A fontFamily string alone loads nothing and every frame renders in a system fallback.
7. Shift the body with <Sequence from={TITLE_LEN}>, never by retiming clicks and pointer keys by hand.

5. Hard rules — each one shipped as a visible flash

These are not style. Each was misdiagnosed at least once before it was understood. Debugging guide: references/render.md.

1. --concurrency=1 on every sequence render. Parallel Chromium returns isolated frames whose page texture is wrapped — a 2×3 grid of the same UI. A handful out of several hundred, so it reads as flicker, and it is random. --gl=swangle does not fix it.

2. No CSS camera on the chrome. Do not wrap the title bar, header or nav in a lasting transform: scale(). This bans moving the frame. It does not ban moving what is inside it — see §3. A click punch is allowed: a short bell returning to exactly 1.

3. Scale footage, never type. A slow push on a video panel is a camera move. The same push on a UI is shimmering text.

4. One layer of any given shot. Hard-cut and unmount rather than fading a video in over an editor already playing it.

5. Dark ground, dark loop — for a page-section clip. Start and end on the same dark field, or the seam strobes on every repeat. The title card makes this free. A standalone clip is exempt from the seam and may end on a card or a logo; the dissolve rule still holds either way — phases cross over the frozen outgoing phase, because fading one out onto white is a brighter flash than the cut.

6. Restart a push at the source's own cut. Find cuts with ffmpeg -vf "select='gt(scene,0.25)'".

   ---

   Rule 5 has a trap, and knowing the rule does not avoid it. Written as above it sounds like advice about easing. It is not: it is about what is underneath.

   The construction you will reach for is to fade the outgoing phase out while fading the incoming one in. At the midpoint both sit near 50%, the ground shows through both, and the seam reads as a dip. Instead, leave the outgoing at full opacity and bring the incoming in over it.

   Then the part that is easy to miss even after doing that: the incoming phase must actually be opaque. A phase that draws panels on a transparent root is a sheet with holes in it — the old phase shows through the gaps, and the frame the old one unmounts on drops all of its panels at once. That is a full-frame luminance jump, check-frames.mjs fails the render, and no amount of easing hides it. Give every phase root its own background.

   And keep the last phase mounted through whatever follows it. If the film holds on a finished state after the flow ends, the phase that drew that state has to still be alive, or the hold is an empty screen.

   All three of these were shipped, in that order, by an agent that had read rule 5 first.

7. One still per timeline frame. Never hold 24fps footage as every-other-frame in a 30fps comp. Extract 1:1.

8. Bump the cache-buster on both src and poster when you replace a clip on a page. A poster without a version sticks.

6. Preview, render, gate

Look at it before you render it

sh scripts/preview.sh <composition-id>

Do this before every full render, and read the contact sheet it writes. It shoots a still on each beat and again 12 frames later, reading the beat sheet straight out of the composition. Seconds, against minutes for a render that hard rule 1 pins to --concurrency=1.

This is not the gate and does not overlap with it. check-frames.mjs measures two numbers per frame — half-against-half similarity, and average luminance. A panel over the header, a phase drawn empty, a counter cut off before it lands, a pointer pressing the number above the button: none of those move either quantity. The gate passes and the clip is still wrong. Every defect in that list has shipped from this repo, each one found by rendering a frame and looking at it.

What to look for, in order: the pointer on the thing it presses (rule 10), every panel inside the frame and clear of the chrome, no phase drawn empty, nothing cut off at a panel edge, and the bottom of the canvas doing work (rule 9).

Then hold a still next to a screenshot of the real screen. Same list, same length, same order? Same grouping and column count? Same ground, light or dark, and the same accent? This is the check for §1's rebuild step, and it is the only one that catches a clip which is beautifully animated and not the product. If nobody can produce a screenshot, that is itself the answer: ask for one before rendering, not after.

What it cannot see is time. A stack of stills says nothing about a rhythm that drags, a beat that lands early, or a hold that outstays its welcome. Those only exist in motion, so the finished mp4 still gets watched.

Render and gate

sh scripts/render.sh <composition-id> [poster-frame]

It renders --sequence --image-format jpeg --jpeg-quality 90 --concurrency=1, stitches to 1536×960, writes the mp4 and poster, runs check-frames.mjs, and fails on its failure.

check-frames.mjs is the regression test for hard rules 1 and 5. Do not "fix" a fail by loosening the detector.

On a failure the frame sequence is kept, and the script prints the path. That is the point of rendering through stills: open the flagged JPEG and see whether the grid is already in it, which is what decides whether the encoder is innocent. references/render.md has the order.

Back up the previous mp4 and poster before re-rendering over them.

7. Sound — ask, once the picture is done

When the render passes, ask whether they want a sound version. Once, in one line, and only then — sound is a separate pass over the finished mp4, so asking earlier just holds up the work.

Which question you ask depends on where the clip lives (question zero).

A clip for a page section. Ask it as where it is going, never as "do you want sound":

The silent clip is done. Autoplaying in a page it has to stay muted — browsers block autoplay with audio. Want me to score a second file for social / docs, where sound plays on tap?

That framing matters here. "Do you want sound?" invites yes, and a yes produces a file that will not autoplay on the page it was made for. The two outputs are not alternatives: the muted mp4 is the deliverable, the scored one is an extra.

A standalone clip. The opposite. Nobody autoplays it, so there is no muting constraint and the scored file is the one that ships — a launch video posted silent looks unfinished. Ask plainly, and expect yes:

Picture is done. Want me to score it? Roughly 30 beats — clicks, panels landing, one whoosh per screen change. It runs over the finished file, so the picture cannot change.

Either way the silent original is never overwritten.

If they say yes:

bash scripts/add-sfx.sh <composition-id>     # out/<id>-sound.mp4

Beats live in scripts/beats/<id>.txt — copy demo.txt and score against the beat-sheet constants at the top of the composition. Frames are written in body coordinates, the same numbers as the constants; the script adds the title-card offset.

The pass copies the video stream byte for byte (-c:v copy), so no frame is redrawn and the gate does not need re-running.

The palette

Nine sounds, and each one is chosen by the level of motion — the DUR scale in src/motion.ts — never by which element it is attached to. Choose by element and the vocabulary grows without bound and the clip stops sounding like one thing.

A single event is always pitch 1.00. Spread of ±6% belongs only inside a burst of identical repeats sitting close together — eight documents, seven belt rotations, a typing run. The machine-gun effect comes from density, not from which sound you picked.

Score events, not motion. A counter counting and a bar growing are motion. Putting a sound on them means a sound on every frame for a second and a half, after which nothing reads as an event. One hero beat per phase applies to sound too.

The defect already made once: whoosh on everything. A whoosh means the screen changed. Score an element arrival and a list rotation with it as well and the entire clip sounds identical.

The shipped sounds are synthesised by ffmpeg on first run — placeholders, and they sound like it. Drop real wavs into out/sfx/ under the same names and every timing still holds; Kenney's UI Audio pack is CC0 and needs no attribution.

8. Reviewing

Remotion has no CSS transitions and prefers-reduced-motion is meaningless in a rendered file. Everything else the web animation literature says about curves, properties, durations and stagger applies here unchanged — §3 is a translation of it, and Emil Kowalski's animate / review-animations skills are the best source to read alongside this one.