Sutra

The reference video is the specification

When a client sends a reference, that file is the brief and the notes around it are pointers into it. The job is to rebuild that film's structure with the brand's own assets, which means counting shots, cuts, sound and type first.

What is in here
  1. What does a client actually mean when they send a reference?
  2. Census the reference before you prompt anything
  3. The failure this prevents is a rebuild of the vibe
  4. How do you know the rebuild is finished?
  5. When is the reference the wrong specification?
  6. How to send a reference so it works
The short answer

When a client sends a reference video, that file is the specification and everything said about it is a pointer into it. The target is not the reference's mood. It is that specific film's compositional structure, rebuilt with the brand's own assets: the same shot count, the same cut positions, the same relationship between sound and picture, the same places type enters and dies. Before anything is prompted or cut, you take a census of the reference, shot by shot, off the file rather than off memory.

What you get out of this
  1. The difference between a mood board, a reference and a verbal note, and what each one can specify
  2. A census template you can copy, with every row a builder actually needs
  3. The failure this prevents: a rebuild of the vibe that passes every check and drops the structure
  4. The acceptance test, where matching the reference is explicitly not good enough
  5. When the reference is the wrong specification, and how to tell before you copy it

Mood board, reference video, verbal note

Compare
Mood board, reference video, verbal note
DimensionMood boardVerbal noteReference video
Specifies a lookYes. Yes, this is what it is forPartly. RoughlyYes. Yes, and in a moving version
Specifies a structureNo. NoNo. NoYes. Completely, once counted
Specifies paceNo. NoNo. Fast and slow mean nothingYes. To the frame
Specifies the soundNo. NoPartly. A genre label at bestYes. Where it starts, what carries it, where it lands
Specifies where type goesNo. NoNo. NoYes. Entry, hold and how it leaves
Can be checked against the delivered fileNo. Only by argumentNo. Only by argumentYes. Side by side, on numbers
All three are useful and only one of them is a specification. Trouble starts when a reference is treated as the first column, because then the builder has to invent the structure and will invent the cheapest one.

01What does a client actually mean when they send a reference?

They mean this one. Not something in this area, not a family of films that feels like this. This one, and usually one specific property of it that they cannot name, because naming it is a video person's job and they are not a video person. Their sentence about it is a pointer, not a description. Follow the pointer into the file.

So the reference outranks the adjectives around it. If somebody says premium and sends a film that cuts sixteen times in fifteen seconds, premium in this brief means that cutting rate, not a slow push on a bottle. If they say calm and send a film with a transient loud enough to move a needle at second three, calm means that. The file wins every argument with the words.

The reference is the specification. Everything said about it is a pointer into it.

the rule, written after we built the adjectives instead of the film

Vote, then see

When you send a studio a reference, what do you actually want copied?

02Census the reference before you prompt anything

A census is a document, produced from the file, that says what the reference is made of. It takes about forty minutes on a fifteen second film and it replaces every later argument about whether the rebuild is close enough. Do it before the first prompt, before the first cut, before the shot list, because a shot list written from a vague memory of a film is a shot list of your own habits.

A reference census, in the order to do it

Walk it
Step 1
Count the shots, mechanically

Scene-detect the file, then watch it and correct the detector by hand. Both halves are necessary. Detectors miss cuts hidden inside motion, and eyes miss cuts that land under a loud sound.

Step 2
Time every cut, in and out

Write the in point, the out point and the duration of each shot. You are looking for the shape as much as the average: whether the film accelerates, decelerates, or sits on one length. An even rhythm is the thing that makes rigorous films read as one film.

Step 3
Name what is in each frame, in five words

Not what it means. What it is. Hand entering empty frame. Product at full frame. Wide of a room with one lit lamp. Five words forces you to name the subject rather than the feeling, and the subject is what you have to find or build.

Step 4
Log the camera, per shot

Locked, push, pull, handheld, whip, or a move that starts inside the shot. This column is where most rebuilds quietly diverge, because a locked frame is harder to source than a drifting one and the drift gets chosen by default.

Step 5
Log the sound against the picture

The frame of first sound, what carries the film, and where the biggest transients land. Then check whether the largest ones sit on picture cuts or on type reveals. In two films we measured they landed on the type, which changes what you are building entirely.

Step 6
Log the type: entry, hold, exit

How long each string is on screen fully opaque, and how it leaves. In one reference we checked to the frame three separate times, every headline vanished on a picture cut and nothing ever faded out. Type appears inside a shot and is killed by the edit.

Step 7
Write three sentences: what is not decoration

Name the three things this film is doing that are load bearing. If you cannot find three, you do not understand the reference yet, and the rebuild will be a costume rather than a structure.

1 / 7

Forty minutes for a fifteen second reference. The output is one table and three sentences, and the three sentences are the part that matters.

A census run, over the reference and over our own source clips

Watch it run
A trimmed transcript, file names changed. The two warnings are the useful part: stock and generated clips are frequently compilations, so a shot you think is one shot can contain a hard cut you did not put there.

The census sheet, as we fill it in

Copy this
REFERENCE CENSUS
file         reference-01.mp4      runtime  14.83s   fps 24   frames 356
cuts         16                    mean shot 0.93s   min 0.21s   max 2.75s
shape        accelerating: blocks shorten through the second half
first sound  0.02s                 mode  music, no speech
brand mark   first at 3.6s (24% of runtime), held 1.1s at full opacity
type         6 strings, every one killed by a cut, none faded out

#   in      out     dur    WHAT IS IN FRAME          CAMERA   TYPE     SOUND
01  0.000   0.417   0.417  hand enters empty frame   locked   -        downbeat
02  0.417   0.958   0.541  product at 100% of frame  push     -        cloth
03  0.958   1.375   0.417  wide, one lit lamp        locked   line 1   -
..  ...     ...     ...    ...                       ...      ...      ...

THREE THINGS THIS FILM IS DOING THAT ARE NOT DECORATION
  1.  it withholds the finished object until 9.2s
  2.  the cut rate halves exactly halfway through
  3.  every headline dies on a cut, so the edit is the punctuation

DEVIATIONS WE INTEND
  what            <the departure>
  why             <what the reference does not cover in our case>
  risk accepted   <and by whom>
One row per shot, filled from the file rather than from memory. The two blocks at the bottom are the ones people skip, and they are the two that make the rebuild defensible when somebody asks why it looks different.

03The failure this prevents is a rebuild of the vibe

Here is how it goes wrong, and it goes wrong quietly. Somebody watches the reference twice, forms an impression, and builds a film that shares the impression: same color temperature, same music genre, same kind of product shot. It passes a language check, because every word in the brief is honored. It ships the cheapest subset of what the reference was actually doing, and nobody can say why it is worse.

What a census looks like laid on a runtime

Second by second
0s1s2s3s4s5s6s7s8s9s10s11s12s13s14s15sFive cuts, object onlySeven cuts, acceleratingThe finished thing, at…End cardLine 1,…Line 2,…The claimEnds loudPICTURETYPESOUND
Read it as a list
AtChannelWhat happens
0.0sPictureHand enters
0.4sPictureFive cuts, object only
3.6sPictureSeven cuts, accelerating
9.2sPictureThe finished thing, at last
12.4sPictureEnd card
1.3sTypeLine 1, killed by a cut
5.1sTypeLine 2, killed by a cut
11.6sTypeThe claim
0.0sSoundSound at frame one
5.0sSoundBiggest transient, on the type
13.0sSoundEnds loud
The shape of one census on a fifteen second reference. Your file's numbers will be different. The rows will not be, and the third row is the one most rebuilds never fill in.

Read the three lanes together and the thing you were supposed to copy becomes obvious. The film withholds its payoff for nine seconds, kills its type with the edit, and puts its loudest moment on a word rather than on a picture cut. A vibe rebuild copies none of those, because none of them are visible to somebody watching twice and forming an impression.

The sound row is the one that gets skipped, because it is the only row you cannot see. A reference that starts on a transient at frame one and one that fades up over half a second are different films, and nothing on a timeline shows you which you are looking at. On a working split of ours, sound is 40% of the ad, and it gets briefed the same way the picture does, off the file.

Four sentences clients say, and what each one is really pointing at

Flip them
None of these is a bad brief. Each one is a pointer, and the job is to follow it into the file rather than to answer the adjective.

04How do you know the rebuild is finished?

You put it beside its parent, at the same size, and ask one question: which of these two would I rather send to a buyer? Not which is closer. Which is more sellable. Until the answer is yours, the work is not done, and matching the reference is explicitly not a pass. The reference library is the baseline, not the target.

Four numbers a census produces, and what they are for

Measured off files, not surveyed
0.83smedian mean shot length across the 48 films we timed, with product films clustering shorterSutra Haus census log
0.018smedian time to first sound; 40 of those 48 films were under 0.12s and none faded inSutra Haus census log
22%of runtime before the product is first named, measured across 68 competitor transcripts, and never in the hookSutra Haus census log
16cuts in one fifteen second reference, which is a description of that film and was a disastrous targetSutra Haus census log
All four are ours, counted off the reference sets we timed rather than taken from a study. They describe what we measured. None of them is a target, and the fourth one is in this list because we tried using it as a target and it made every film the same.

That last one is worth the warning. We once turned a census into a spec sheet, wrote fourteen to eighteen cuts and seventeen to nineteen seconds into the doctrine, and got twenty films that a viewer would call one ad made twenty times. A census describes a film. It does not prescribe your film, and the numbers in it are evidence rather than instructions. There is more on that in duration is the action's pace, not a length.

05When is the reference the wrong specification?

Three cases, and all three are common enough to check for by default. The first is borrowed grammar: a device that carries meaning inside the reference brand's own system and carries nothing outside it. The color flash is ours, and we shipped it before we understood it. Distinctive assets work because one brand owns them, which is exactly why they do not travel.

The second is the brand gap. A famous brand's film can open on mood because the viewer already knows what the product is. An unknown brand copying that form inherits the mood and none of the recognition, so the film is beautiful and answers nothing. The third is coverage: a reference can specify a shot your brand has no way to produce, and the honest response is a refusal that names what would unlock it.

Four films, four different references, counted first

Ours, four different parents
Ephoria - campaign ad
Hotel client - campaign reel
SaaS - studio demo
Cafe - studio demo
Each of these started as somebody else's film on a table and a census sheet beside it. If you can tell which two came from the same studio, we lost the argument in section five.
One thing that is not a problemCopying a structure is not a legal risk. A cut pattern, a parallax rig, a kinetic type system are all methods, and a procedure or method of operation is outside copyright. The risk in copying a reference is never the law. It is copying a device that only worked because of who owned it. That is also why one visual language per brand is a production rule here rather than a slogan.

Four first frames, four different structures

Four openings
Ours, from four different brands. Each opening was chosen from a census of a different reference, which is why they have nothing in common. If they had something in common, that would be our fingerprint rather than the brands' languages.

06How to send a reference so it works

One link, one sentence, one exclusion. The link is the file, not a playlist. The sentence names the structural property you mean, in your own words: it holds one object for four seconds, or it never shows a face, or every line of text disappears on a cut. The exclusion names the thing you do not want carried over, which is usually the brand's own signature device.

If you send three references, rank them and say what each one is for, because three unranked references specify nothing and force somebody to average them. Averaging three good films produces a fourth film with no argument, which is the quiet way a brief becomes generic. The rest of the brief is covered in how to brief and QA ad creative.

What to send with a reference, and what to say about it

Tick as you go - it remembers
0%
Five minutes on your side. It removes the single most common cause of a second round, which is a builder guessing at which property of your reference you actually meant.

Questions people actually ask

Open what you need
What does a reference video mean in a creative brief?

It means that film specifically, and it functions as the specification rather than as inspiration. Everything said around it is a pointer into the file. The usable version is the structure: shot count, cut positions, camera per shot, where sound starts and where its biggest moments land, and where type enters and how it leaves.

What is the difference between a mood board and a reference video?

A mood board specifies a look and cannot specify anything that happens over time. A reference video specifies pace, structure, sound and type placement to the frame, once somebody counts them.

The practical difference: a mood board can only be argued about, and a reference can be checked side by side against the delivered file on numbers.

How do I use a reference video without copying the ad?

Copy the structure and change the argument. The shot count, the cut rhythm and the sound relationship are methods, and methods are not what makes an ad recognizable. What makes it recognizable is the subject, the world, the point of view and the brand's own devices, and those are exactly the things you should not carry over.

How many references should I send?

One, ideally. If you send three, rank them and say what each one is for, because three unranked references get averaged and the average of three good films is a film with no argument. One reference plus one sentence about what you mean beats a folder every time.

What if the studio says the reference cannot be matched?

That is a good answer if it names what would unlock it. A refusal that says we cannot hold that four second shot because we have no footage of the product in daylight, send one clip and it becomes possible, is a ten-minute job for you. A refusal with no cure attached is a studio that has stopped looking.

Should the finished ad look like the reference?

It should share the reference's structure and not its identity. Put the two side by side at the same size and ask which you would rather send to a buyer. If the honest answer is the reference, the work is not finished, because matching a baseline is not the same as beating it.

A reference sent without a census is a wish. A reference counted, row by row, is a contract that both sides can check against the delivered file, which is the only kind of brief that ever settles an argument before the invoice.

Where the numbers came from

  1. Motion. Creative Benchmarks, 550,000+ ads across 6,000+ advertisers - used for the point that distinct swings, not variations of one film, are what move an account
  2. Ehrenberg-Bass Institute for Marketing Science. Institute research on distinctive brand assets - why a device that carries meaning inside one brand carries none inside another
  3. United States Copyright Office. Circular 33: Works Not Protected by Copyright - a procedure, process, system or method of operation is not protectable, which includes a cut pattern

Every figure above links to the place it was published. Numbers marked as ours are measured inside this studio and we say so where they appear. We do not print a statistic we cannot point at.

Badal Kariwal

Runs Sutra Haus, a one-person ad studio that has shipped over a thousand finished creatives - film and stills - for DTC brands and hotels. Writes here about what the work actually taught him, including the parts that failed. The person who reads your brief is the person who builds the work. Send him something to make.

Send the reference, get the census back

Send a link and the film you wish yours looked like.

We build one finished ad from your own product, free, inside three days, yours to run whether or not we ever work together. You also get the census of your reference: shots, cuts, sound and type, counted off the file. Keep it and hand it to whoever builds your next one.

Replies within a day. Ad within three.
Read next