Script

Write script

This script is the single source for everything that follows: shoot plan, cut list, subtitles, and caption are derived from it. That's why the output format is binding and must not vary.

Process

1. Load context. BRAND.md for the voice, FORMATS.md for the structure of the chosen format, LEARNINGS.md for what's worked most recently.
2. Clarify the framework. Platform, target length, pillar, and format. Derive whatever wasn't specified from the project knowledge and make the assumption visible above the script — don't ask, as long as a reasonable assumption is possible.
3. Set the promise. One sentence: what does the viewer get out of it after 30 seconds? Sits as a comment line above the script and is the touchstone for every beat.
4. Build beats. Every beat carries exactly one thought. A beat that wants two things at once gets split.
5. Calculate speaking duration (see below) and derive the timecodes from it — don't estimate.
6. Run the self-check and output the result below it.

Speaking duration

Calculate with 2.7 spoken words per second. That's a brisk pace, typical for social video.

• Count the word count per beat, divide by 2.7, round to half seconds.
• Add 0.3 seconds of buffer per beat for breathing and cutting.
• Set beats without a spoken line (pure visual moments) at the planned hold time, at least 0.8 seconds.

If the total deviates from the target length by more than 15 percent, shorten or extend the content — never by speaking faster.

Calibration: after the first three videos are shot, compare the actual speaking duration to the calculated one and correct the value 2.7 in this file to your own pace. After that, the lengths will be accurate to the second.

Output format

Exactly this structure, with no additional headings or body text in between:

# SCRIPT · <working title>
PROMISE: <one sentence>
META: platform=<reel|tiktok> · target=<seconds>s · calculated=<seconds>s · pillar=<name> · format=<name>

## BEAT 1 · HOOK · 00:00–00:02
VO:    "<spoken text, verbatim>"
TEXT:  <overlay, in caps as it should appear, or ->
SHOT:  <framing, movement, what's visible>
SOUND: <music change, effect, or ->

## BEAT 2 · <LABEL> · 00:02–00:09
VO:    "..."
TEXT:  ...
SHOT:  ...
SOUND: ...

## BEAT n · CTA · 00:xx–00:xx
VO:    "..."
TEXT:  ...
SHOT:  ...
SOUND: ...

Field rules:

• VO is verbatim and speakable. Read it aloud in your head: what can't be said in one breath gets split. No written-register phrasing, no nested clauses, no words that trip you up.
• TEXT is the on-screen overlay. Short, in caps, never identical to what's spoken – it complements or anchors it. Empty is better than redundant, then -.
• SHOT describes framing, movement, and subject concretely enough to shoot from without having to ask.
• SOUND only gets an entry when something happens at that point: music change, beat drop, effect, silence. Otherwise -.
• LABEL names the beat's function: HOOK, PROBLEM, MISTAKE 1, TWIST, PROOF, RESOLUTION, CTA.

Format-specific fields

If FORMATS.md describes a text carrier in the shot – a sign, a card, a printed document, an object with lettering –, then that's a field of its own in the beat, not an aside under SHOT:

SIGN: <exact text, as it should appear on the carrier>

The text has to fit the surface and work without sound. A script that only describes it instead of spelling it out forces reinvention later.

Check in FORMATS.md whether the carrier exists on set or only gets created in editing. A physical prop has to be made before the shoot; an overlaid card is created in post and can be changed until the last moment. Don't treat it as a shoot task if it isn't one.

The same applies to any other recurring element that FORMATS.md names as part of the base format. Carry it over into the schema as a named field.

Self-check

Output after the script, each marked met or not:

• Hook under twelve words and with its own visual idea
• All format-specific fields from FORMATS.md are filled in, not paraphrased
• Every beat carries exactly one thought
• No TEXT field repeats its VO field verbatim
• Calculated length within 15 percent of the target
• The promise is actually delivered by the end
• The CTA matches the promise and is a single, concrete action
• Sounds like BRAND.md – no phrasing that could sit on just any account
• If FORMATS.md defines a check formula, it's satisfied point by point – it beats all the general rules in this instruction

Whatever isn't met gets named and fixed, not glossed over.