Process
-
Lock inputs and route.
- Premise options requested: call the Skill tool with
"ad-concepts"and stop. - Existing draft script supplied: read
draft-audit.mdand run that workflow instead of steps 2 to 5. - New script requested: ask once only if product or exact runtime is missing; when a direct brief omits a premise ("Write a 20-second ad for X"), commit to one dramatic situation passing the swap test and proceed.
- Exit gate: four inputs are locked before step 2: exact runtime in seconds, product with one single-minded proposition (secondary claims parked out of scope), one product-locked dramatic situation, and any tone or transferable mechanisms.
- Premise options requested: call the Skill tool with
-
Build the timed visual spine first.
- Populate
TimeandVideofrom opening hook through physical proof and brand entry to closing button before writing any spoken line. - Express every emotion and product claim as camera-observable reality (physical actions, objects, facial reactions, environment, diegetic text).
- Exit gate: contiguous
MM:SS-MM:SSbeats sum strictly to target runtime with zero gaps or overlaps, everyVideocell passes the camera-lens test (zero internal states, off-screen backstory, or abstract summaries), and physical proof plus brand entry land on or before the button.
- Populate
-
Layer Audio as counterpoint.
- Populate
Audio(## Format) as audio counterpoint: every spoken line adds subtext, irony, rhythm, or a twist absent fromVideo. Replace any line that captions visible action withSFX:or an earned pause ([PAUSE Ns]). - Place at least one earned pause or visual/SFX-only beat at the key reveal, reaction, or proof turn (omit only for declared patter).
- Exit gate: zero captioning lines remain, and every silence or reaction hold carries an explicit
[PAUSE Ns]duration.
- Populate
-
Audit the read-aloud budget.
- Pick a delivery tier from
## Cadenceand verify spoken pace per beat and across the full spot:- Count spoken words only (
VO:andCHARACTER:); exclude cue labels,SFX:,MUSIC:, and parenthetical delivery notes. - Subtract every
[PAUSE Ns]and visual/SFX-only hold from beat duration to get net speaking seconds. - Check both words/sec and syllable density against the declared tier.
- Count spoken words only (
- When any beat or total exceeds ceiling, cut spoken words while preserving earned pauses; adjust beat durations only when visual action needs more room.
- Exit gate: every beat and the full spot sit at or under the stated cadence ceiling, all earned pauses remain intact, and cumulative timestamps still match target runtime.
- Pick a delivery tier from
-
Run the mute test.
- Strip all spoken audio (
VO:andCHARACTER:) and explanatory supers or cards. - Perform a blind read strictly of
Videoas written: state what a muted viewer learns about the product, cite the exact beat timecode(s) carrying physical proof, and separately list what lives only inAudio. - Exit gate: the single-minded proposition is provable strictly from the cited
Videobeats (Verdict: PASS). If any part of the core claim depends on spoken audio or explanatory copy, rewriteVideoso physical action carries the proof, then re-run step 4 before delivering.
- Strip all spoken audio (
Cadence
- Baseline conversational:
<= 2.3 words/secin English,<= 2.0 words/secin longer-word languages (Russian, German, etc.);<= 4.5 syllables/secbaseline ceiling across languages. - Deadpan or dramatic:
<= 1.8 words/sec(EN) /<= 1.6 words/sec(longer-word languages). - Deliberate patter: up to
3.0 words/sec(EN) /2.6 words/sec(longer-word languages); declare explicitly in Header.
Format
- Header: Product, single-minded proposition, exact runtime, brand entry timestamp, delivery cadence and wps ceiling.
- Three-column script (Markdown table):
| Time | Video | Audio |
|---|---|---|
0:00-0:03 | Camera-observable action | VO:, CHARACTER:, SFX:, MUSIC:, [PAUSE Ns] |
- Read-aloud budget: per-beat lines (
0:00-0:03: 4w / 2.0s = 2.0 wps, [PAUSE 1s]or[visual/SFX only]) plus totals (spoken words,pause/visual-only sec,net speaking sec,overall wps,Verdict: PASS). - Mute test:
Visual-only read(citing proof beat timecodes),Audio-only layer,Verdict: PASS.