Studio · Edit · 9:16 · cutdowns
Retention Edit
A retention-editing line: analyze a reference into cuts, beats, energy curve and drop; rebuild the structure with your own footage; land the hero on the drop and export platform ratios. The structure is copied, the footage never is, and every cut is a decision with a reason.
The whole line, one pack
story to stitch
The board below is the core, not a ceiling; extend it until the story lands.
What you get
11 deliverables
- Reference analysis: cut map, beat map, BPM, energy curve, drop timestamp
- Written blueprint: structure only, with what is explicitly not copied
- Cut plan over the transcript: trim, split, concat, reorder
- Beat markers and one clip per shot at the same start and duration
- Hero shot landed on the drop
- Word-level captions and a consistent grade
- Exports at 9:16, 1:1 and 16:9
- Retention notes: what was tested and what the numbers said
- Beat grid file with BPM, downbeats and the drop marker
- Cut log: every cut with its timecode, its beat and its reason
- Retention read at 3 s, 10 s and the drop, per platform
World and look
paste before anything
| World | YOUR REFERENCE - only a video you have the right to use as a reference. Never reproduce its actual footage. The world of the rebuild is your own footage, your own subject and your own voice. |
| Palette | YOUR PALETTE - one grade for the whole cut, chosen before the first edit. Captions in one font and one weight. No look borrowed from the reference beyond structure. |
| Negative | extra fingers, fused hands, face morph, changing eye colour, plastic skin, warped jewellery, floating earrings, captions, watermark, logo, extra teeth |
Locks
location, props, wardrobe, axis
| rights | A written note of the right to study the reference: owned, licensed, or publicly available for analysis. No right, no teardown. |
| footage | Your own clips, logged with duration, resolution and content. The reference footage is never imported into the timeline. |
| beats | The BPM and the beat grid are locked before the edit. Every cut either lands on a beat or deliberately breaks one. |
| canvas | One canvas per export: 9:16, 1:1, 16:9. Frame rate and resolution match the source footage, never upscaled in the edit. |
| captions | One font, one weight, one position. Captions are word-level and never cover the subject's mouth. |
| transcript | One transcript, corrected by hand before the cut plan. Auto-captions are a first draft; a wrong word in a caption is a wrong word in the edit. |
| export | One export preset per ratio, saved and reused. Bitrate, frame rate and audio level are part of the lock, not a per-video decision. |
| subject | One subject per cut: the same face, hair and garment across every source clip. If the footage has two looks, the cut has two videos and one of them is wrong. |
| loop | The last frame matches the first within two pixels, with no fade. A loop that jumps is a replay the platform counts and the viewer resents. |
| energy | The rebuilt energy curve is drawn over the reference's before export. A flat stretch longer than four seconds is a cut problem, not a music problem. |
Every lock is pasted into the prompt for every shot that touches it. A lock that is paraphrased is not a lock.
Cast locks
0 identities
No cast. This pack is product, archive or graphic-led.
Shot prompt grammar
16 rules every prompt on this board obeys
- Order: Subject, Action, Camera, Style, Constraints (S-A-C-S-C). Subject first so the model binds the identity before it reads motion.
- Actions stay slow, soft and continuous. One main camera move per shot plus one texture cue (rain on glass, dust in a beam, fabric settling). Two moves in one prompt cancel each other.
- 50-80 words, three sentences, one visual promise. "Cinematic" is not a quality switch: name the light, the lens and the stock instead.
- Positive constraints only: say what is in frame, never list what must not be there. Negatives are reserved for subtitles and audio ("no captions", "no music bed") - the one place a model reads them reliably.
- Foreground on every shot: an out-of-focus object across roughly a third of the frame, between the lens and the subject (see the Foreground line on each shot). Say "blurred" or the model renders it sharp. Behind it, one atmospheric layer - fog, dust, steam or a volumetric beam - so the frame has depth instead of a flat backdrop.
- Length matches the file. A 30s plan runs 0-5s hook, 5-16s expand, 16-25s proof, 25-30s close, written as second-level segments (0s-3s: ...). If the timeline does not match the requested duration the model drops the ending. On a model that ignores seconds, fall back to shot numbers; on 2.5-class models seconds are first-class.
- Chain with first/last frames: the last frame of clip N becomes the first frame of N+1. Generate that handoff still first, check it against the cast lock, then generate the clip.
- One job per reference image: one carries the face, another the wardrobe, another the product, and each declares what it must not adopt. Name the job in the prompt, not only on the image - "strictly follow clip 1's action and camera" is enough, and re-describing a reference frame by frame introduces conflict. Stay inside the model cap (Seedance 2.5: 30 images, 10 videos, 10 audio).
- Audio line uses one symbol set - () music, <> sound effects, {} dialogue, [] on-screen text - and carries one promise, not a bed of everything.
- Cast gets a sheet before a single shot: front, side, half, close-up, generated once and attached to every prompt. Re-describing the character each shot is the largest single cause of wasted renders.
- Performance is written, not labelled: name the limb and the magnitude ("slowly raises her hand", "turns fast, breathing hard"), bridge one action into the next ("using the turn's momentum she raises her hand"), and play emotion as body detail (eyes welling, fingers tightening on the hem) instead of naming the feeling. Avoid sprints, jumps and rolls inside one shot - hand the energy to the cut and the camera.
- Light gets its own sentence: direction, quality (hard or soft), contrast. Colour runs in three layers - a base tone for the whole frame, a mid tone tied to the subject, one or two accent highlights that repeat. Then delete the word cinematic.
- Close every long prompt with the rules paragraph: the consistency lock, no captions or watermarks, and the tension arc that must survive to the last frame. It is the red line the model does not cross.
- Debug at draft resolution. A 480p draft pass is billed at draft rate and answers the question a full-resolution retry would charge full price for; draft on the cheap fast tier where one exists (Seedance 2.5 Flash, opened 2026-11-15: 480p/720p at 46 yuan per million tokens against 70 for 2.5, about twice as fast), and upgrade only the winner. Retakes eat most of a budget - roughly seven of ten credits die on a bad take - so previs the complex shot as a white model first (7-10 retries down to 2-3), shoot 3-5s pieces and stitch, and change one variable per retry.
- Gate the first frame before you scale: wrong face or product, garbled text, malformed hands, CGI look. A failed render gets diagnosed in concrete words first - material, light path, aerial perspective, overlapping structure, a negative that did not hold - then repaired with those words. A blind re-roll with a new seed doubles the bill; only after the repair fails do you change the seed.
- Read the reference post before you write: pull its transcript and overview frame sheet locally so line timing and cut rhythm come from the real post, not a guess.
The same block is baked into the SKILL.md: every shot below is written in this grammar, carries its own foreground line, and the agent keeps the grammar when it adds shots.
Skills this pack loads
3 skills, in order
| Skill | Job in this run |
|---|---|
video | shots, motion, editing and the cut |
social | organic and platform format |
copy-editing | copy and caption revision |
The SKILL.md installs these in order. Editing packs load video, social and copy revision; creation packs load image, video, copywriting and the subavatar. A pack never loads a skill it does not use.
Ad sources this pack reads
plug in your library
| Source | What it gives |
|---|---|
| Foreplay | swipe file, ad copy, video downloads and timestamps |
| Meta Ad Library | live ads, copy, start dates and platforms |
| TikTok Creative Center | top ads, hooks, music and formats |
| AISLE shelf | measured posts with cut manifests and frame sheets |
| Paste, CSV or screenshot | any ad the buyer can see and has the right to study |
If the buyer names a library, the agent uses its export or API. If not, it asks for one export, URL, screenshot, transcript or CSV. Every source returns the same record, so the board and the render contract do not change.
Any model, any seat
swap on a rate limit
Claude, ChatGPT, DeepSeek, Grok, Kimi, Muse, Gemini, Qwen and Seedance can each hold any seat: read the source, write the board, hold the brand voice, run QC. On a rate limit or outage the agent saves the run state, switches model and continues from the last completed step. It never restarts the run and never re-charges a finished take. The core board scales: ten cast members or a hundred and twenty shots are allowed when the script needs them.
The shot board
12 core shots · full prompts
0.0-2.0s
-
hard cut
Hook — A pattern breaks in the first two seconds. Start mid-action.
Edit instruction: open on the strongest frame, no fade, no title. Cut on the first beat.
foregrounda soft foreground shape thrown out of focus across the bottom third, out of focus, between the lens and the subject
audioMusic starts on the cut.
on screen[hook, 5 words]
2.0-4.5s
-
beat cut
Promise — The promise is stated in one line over the second-strongest frame.
Edit instruction: one clip, one line, cut on beat two.
foregrounda blurred doorframe edge across the left third, out of focus, between the lens and the subject
audioMusic bed under voice.
on screen[promise]
line[the promise]
4.5-7.0s
-
beat cut
Build one — Establish the pattern: same framing, same rhythm, third clip.
Edit instruction: repeat the framing of shot 02 so the break later reads.
foregroundthe out-of-focus edge of a near wall across the right third, out of focus, between the lens and the subject
audioBed continues.
7.0-9.5s
-
beat cut
Build two — The pattern holds; one detail changes.
Edit instruction: same shot size, new subject detail, cut on beat four.
foregrounda blurred doorframe edge across the left third, out of focus, between the lens and the subject
audioOne new element enters.
9.5-12.0s
-
speed ramp
Proof — A visual proof: the thing working, in one unbroken shot.
Edit instruction: speed ramp into the proof, hold the result for 12 frames.
foregroundthe out-of-focus shoulder of a chair back across the right third, out of focus, between the lens and the subject
audioEffect sound on the result.
on screen[the proof]
line[what it proves]
12.0-14.5s
-
match cut
Escalate one — A shape match carries the eye into a new location.
Edit instruction: match the shape, not the subject. Cut on the match.
foregrounda blurred strand of hair across the left third, out of focus, between the lens and the subject
audioMusic lifts.
14.5-17.0s
-
whip
Escalate two — Shorter shots, faster cuts, one new element per cut.
Edit instruction: three clips in 2.5 s, each on a beat subdivision.
foregrounda blurred doorframe edge across the left third, out of focus, between the lens and the subject
audioPercussion enters.
17.0-19.5s
-
hard cut
Turn — The obvious answer is wrong. Withhold the reveal.
Edit instruction: cut to black for 4 frames, then the wrong answer.
foregrounda hanging fabric edge, blurred across the top third, out of focus, between the lens and the subject
audioMusic drops out for 4 frames.
line[the wrong answer]
19.5-22.0s
-
beat cut
Reveal — The real answer lands on the beat.
Edit instruction: reveal on the downbeat, no transition.
foregrounda blurred doorframe edge across the left third, out of focus, between the lens and the subject
audioDrop hits.
on screen[the reveal]
line[the real answer]
22.0-24.5s
-
slow motion
Payoff — The hero moment, held long enough to feel earned.
Edit instruction: slow to 40%, hold 1.5 s, then back to speed.
foregrounda hanging fabric edge, blurred across the top third, out of focus, between the lens and the subject
audioMusic peaks.
24.5-27.0s
-
hard cut
Button — The line the audience repeats.
Edit instruction: one clip, one line, cut on the last beat.
foregrounda soft foreground shape thrown out of focus across the bottom third, out of focus, between the lens and the subject
audioMusic button.
on screen[the button line]
line[the button line]
27.0-30.0s
-
loop cut
Loop — The last frame matches the first so the replay is seamless.
Edit instruction: match the first frame within 2 px; no fade.
foregrounda hanging fabric edge, blurred across the top third, out of focus, between the lens and the subject
audioTail that loops into the first beat.
Core board, not a cap. Add shots, scenes and AI-generated cast until the cut works; keep the face, wardrobe and light locked across every new shot.
Beat sheet
8 beats
| Time | Beat | Job |
|---|---|---|
| 0-5% | Hook | A pattern break in the first two seconds. No logo, no intro. |
| 5-15% | Promise | One line that says what the viewer gets if they stay. |
| 15-30% | Build | Establish the pattern the cut will break later. |
| 30-45% | Proof | A visual proof, not a claim. One detail the eye can check. |
| 45-60% | Escalate | Energy climbs: shorter shots, faster cuts, one new element. |
| 60-75% | Turn | The obvious answer is wrong. The cut withholds the reveal. |
| 75-90% | Drop | The hero shot lands exactly on the drop. |
| 90-100% | Button | The line the audience repeats, then a loop cut back to the first frame. |
Moodboard prompts
15 images
Hooks and variants
9 openers
- Hook-first cut: open on the payoff and run the build as the explanation.
- Silent cut: captions and effects only, for autoplay.
- Loop cut: the last frame is the first frame, built for replays.
- Three-beat cut: 15 seconds, one proof, one reveal, one button.
- Tutorial cut: same structure, slower build, more proof.
- B-roll cut: no face, product and hands only.
- Reverse cut: open on the payoff and run the build as the explanation.
- Text-only cut: no voice, captions carry the whole script, built for muted feeds.
- No-drop cut: the reveal lands on a hard stop instead of a drop, for a track that has no drop.
Sound and music
direction, not a playlist
The music is the clock. Mark the BPM and the beat grid before the first cut, put the drop on a marker, and let every cut either land on a beat or break one on purpose. Keep voice above the bed, duck 10 dB under speech, and leave four frames of silence before the reveal. Export with the music at -14 LUFS and true peak -1 dB. Treat the music as the timing authority: mark the BPM, the downbeats and the drop, then place picture against those marks. Duck the bed 10 dB under speech, leave four frames of silence before the reveal, and put the loudest moment of the mix on the drop. If a cut has no beat and no reason, delete it.
QC gate
16 checks, pass or fail
| Check | Pass | Fail |
|---|---|---|
| Rights | The right to study the reference is recorded and the footage is yours | No right: stop |
| Structure | The blueprint copies structure only, not look, subject or performance | A copied frame: rebuild it |
| Beats | Every cut lands on a beat or a deliberate break | A random cut: re-time |
| Drop | The hero shot lands on the drop within 2 frames | A late drop: shift the clip |
| Hook | The first two seconds break a pattern | A slow open: re-cut the first shot |
| Captions | Word-level, legible at 390 px, clear of the mouth | Overlap: move the block |
| Grade | One grade for the whole cut | A shot out of family: re-grade |
| Audio | Voice above bed, no clipping, silence before the reveal | A muddy mix: rebuild |
| Ratios | 9:16, 1:1 and 16:9 are re-framed, not cropped | A lost subject: re-frame |
| Disclosure | Reference credited and AI label where synthetic | Missing: block delivery |
| Transcript | Captions are corrected against the script, not the auto-transcript | A wrong word: rebuild the caption layer |
| Preset | Every export uses the saved preset for its ratio | A one-off setting: re-export |
| Retention | The 3 s, 10 s and drop reads are recorded per platform | No read: the cut is not measured |
| Pattern | The pattern is established before it is broken | A break with no pattern: add the establishing shots |
| Silence | Four frames of silence sit before the reveal | A busy reveal: cut the bed for four frames |
| Loop offset | The last frame matches the first within two pixels | A visible jump: re-cut the loop frame |
What would make this fail
12 known failures
| Symptom | Cause | Fix |
|---|---|---|
| Viewers drop at 3 seconds | The hook explains instead of breaking a pattern | Open mid-action and cut the first line |
| Cut feels busy | Too many transitions, no pattern | Establish one pattern and break it once |
| Reveal does not land | The drop was not marked before the edit | Mark the drop and re-time the last four shots |
| Rebuild feels like a copy | Look and subject were copied with the structure | Replace the footage and the grade, keep the beat map |
| Captions hurt retention | They cover the subject or lag | Move and re-time, word-level |
| Export looks soft | Footage was upscaled in the edit | Match the canvas to the source |
| Audio clips | Bed and voice summed without headroom | Re-mix with a limiter |
| Replay does not happen | The last frame does not match the first | Match the first frame within 2 px and cut the fade |
| Captions carry errors | Auto-captions shipped unedited | Correct against the script and re-export |
| Exports look different per platform | Each export used its own settings | Apply the saved preset per ratio |
| Cut feels random | No pattern was established before the break | Show the pattern twice before breaking it |
| The rebuild copies the reference | Its grade, subject or performance travelled with the structure | Replace all three and keep only the beat map |
Agent runbook
19 steps
Rebuild a proven structure with your own footage: analyze the reference into cuts and beats, write the blueprint, cut on the beat grid, land the hero on the drop, and export every ratio. The structure is the asset; the footage never is.
Confirm the right to study the reference and record it.
Analyze the reference: cut times, shot lengths, BPM, energy curve, drop timestamp, caption style.
Write the blueprint as structure only: hook, promise, build, proof, escalate, turn, drop, button.
State plainly what is not copied: look, subject, performance, script.
Log your own footage with duration, resolution and content.
Write the cut plan over the transcript: trim, split, concat, reorder, with a reason per cut.
Mark the beat grid and the drop on the timeline before placing clips.
Place one clip per shot at the same start and duration as the blueprint.
Reopen the cut at 10%, 35%, 65% and 90% and reject boring or repetitive sequences.
Land the hero shot on the drop and check it within two frames.
Add word captions and one grade; check at 390 px.
Export 9:16, 1:1 and 16:9 and write the retention notes.
Mark the BPM, downbeats and drop before placing a single clip, and keep the grid visible for the whole edit.
Correct the transcript by hand against the script before writing the cut plan.
Keep a cut log with a reason per cut; a cut with no beat and no reason is deleted.
Export every ratio from the saved preset and record the retention read at 3 s, 10 s and the drop.
Test three first frames at thumbnail size before committing to the hook.
Draw the rebuilt energy curve over the reference's before the final export and fix any flat stretch.
Measure the loop offset in pixels before export rather than eyeballing the join.
The prompt
paste into Claude or Codex
You are the production agent for the AISLE playbook "Retention Edit".
MISSION
Rebuild a proven structure with your own footage: analyze the reference into cuts and beats, write the blueprint, cut on the beat grid, land the hero on the drop, and export every ratio. The structure is the asset; the footage never is.
INPUTS
- brief (required): Product, script or reference and the single promise
- assets (optional): Your own cast, product or footage (or generate original)
- runtime (optional): Target runtime and output ratios. Example: 9:16, 15s
- cap (optional): Hard spend cap for this run
Ask once for any missing required input. Do not invent a product, a face or a claim.
PRODUCTION LINE (run in this order; this is the whole job)
01. Source audit: Log every source clip with duration, resolution, rights and the moment worth clipping. Record the URL and the date.
02. Reference read: Measure the reference cuts: in/out times, shot lengths, BPM, energy curve, caption style. Record numbers, not adjectives.
03. Blueprint: Write structure only: hook, promise, build, proof, escalate, turn, drop, button. State what is not copied: look, subject, performance, script.
04. Clip plan: One row per clip: source in/out, recipe, hook line, caption, export ratio. Cut the moment, not the whole video.
05. Cut: Apply the ten recipes to the source. No generated face, voice or line; only what exists in the footage.
06. Captions: Word-timed captions, one accent, safe area checked. No mid-word cuts, no captions over the face.
07. Sound: Loudness to platform spec, music ducked under speech, no watermark and no borrowed audio.
08. Batch and deliver: Export 9:16, 1:1 and 16:9 as needed, named by source and recipe. The batch is the product.
Receipt for every export: source, in/out, recipe, platform. No model takes to price; the edit is the spend. The source is the only footage.
WORLD
YOUR REFERENCE - only a video you have the right to use as a reference. Never reproduce its actual footage. The world of the rebuild is your own footage, your own subject and your own voice.
Palette: YOUR PALETTE - one grade for the whole cut, chosen before the first edit. Captions in one font and one weight. No look borrowed from the reference beyond structure.
Negative prompt: extra fingers, fused hands, face morph, changing eye colour, plastic skin, warped jewellery, floating earrings, captions, watermark, logo, extra teeth
CAST LOCKS (paste the full lock into every prompt)
- No cast. This is a product, archive or graphic-led film.
LOCKS (location, props, wardrobe, axis)
- rights: A written note of the right to study the reference: owned, licensed, or publicly available for analysis. No right, no teardown.
- footage: Your own clips, logged with duration, resolution and content. The reference footage is never imported into the timeline.
- beats: The BPM and the beat grid are locked before the edit. Every cut either lands on a beat or deliberately breaks one.
- canvas: One canvas per export: 9:16, 1:1, 16:9. Frame rate and resolution match the source footage, never upscaled in the edit.
- captions: One font, one weight, one position. Captions are word-level and never cover the subject's mouth.
- transcript: One transcript, corrected by hand before the cut plan. Auto-captions are a first draft; a wrong word in a caption is a wrong word in the edit.
- export: One export preset per ratio, saved and reused. Bitrate, frame rate and audio level are part of the lock, not a per-video decision.
- subject: One subject per cut: the same face, hair and garment across every source clip. If the footage has two looks, the cut has two videos and one of them is wrong.
- loop: The last frame matches the first within two pixels, with no fade. A loop that jumps is a replay the platform counts and the viewer resents.
- energy: The rebuilt energy curve is drawn over the reference's before export. A flat stretch longer than four seconds is a cut problem, not a music problem.
SHOT PROMPT GRAMMAR
- Order: Subject, Action, Camera, Style, Constraints (S-A-C-S-C). Subject first so the model binds the identity before it reads motion.
- Actions stay slow, soft and continuous. One main camera move per shot plus one texture cue (rain on glass, dust in a beam, fabric settling). Two moves in one prompt cancel each other.
- 50-80 words, three sentences, one visual promise. "Cinematic" is not a quality switch: name the light, the lens and the stock instead.
- Positive constraints only: say what is in frame, never list what must not be there. Negatives are reserved for subtitles and audio ("no captions", "no music bed") - the one place a model reads them reliably.
- Foreground on every shot: an out-of-focus object across roughly a third of the frame, between the lens and the subject (see the Foreground line on each shot). Say "blurred" or the model renders it sharp. Behind it, one atmospheric layer - fog, dust, steam or a volumetric beam - so the frame has depth instead of a flat backdrop.
- Length matches the file. A 30s plan runs 0-5s hook, 5-16s expand, 16-25s proof, 25-30s close, written as second-level segments (0s-3s: ...). If the timeline does not match the requested duration the model drops the ending. On a model that ignores seconds, fall back to shot numbers; on 2.5-class models seconds are first-class.
- Chain with first/last frames: the last frame of clip N becomes the first frame of N+1. Generate that handoff still first, check it against the cast lock, then generate the clip.
- One job per reference image: one carries the face, another the wardrobe, another the product, and each declares what it must not adopt. Name the job in the prompt, not only on the image - "strictly follow clip 1's action and camera" is enough, and re-describing a reference frame by frame introduces conflict. Stay inside the model cap (Seedance 2.5: 30 images, 10 videos, 10 audio).
- Audio line uses one symbol set - () music, <> sound effects, {} dialogue, [] on-screen text - and carries one promise, not a bed of everything.
- Cast gets a sheet before a single shot: front, side, half, close-up, generated once and attached to every prompt. Re-describing the character each shot is the largest single cause of wasted renders.
- Performance is written, not labelled: name the limb and the magnitude ("slowly raises her hand", "turns fast, breathing hard"), bridge one action into the next ("using the turn's momentum she raises her hand"), and play emotion as body detail (eyes welling, fingers tightening on the hem) instead of naming the feeling. Avoid sprints, jumps and rolls inside one shot - hand the energy to the cut and the camera.
- Light gets its own sentence: direction, quality (hard or soft), contrast. Colour runs in three layers - a base tone for the whole frame, a mid tone tied to the subject, one or two accent highlights that repeat. Then delete the word cinematic.
- Close every long prompt with the rules paragraph: the consistency lock, no captions or watermarks, and the tension arc that must survive to the last frame. It is the red line the model does not cross.
- Debug at draft resolution. A 480p draft pass is billed at draft rate and answers the question a full-resolution retry would charge full price for; draft on the cheap fast tier where one exists (Seedance 2.5 Flash, opened 2026-11-15: 480p/720p at 46 yuan per million tokens against 70 for 2.5, about twice as fast), and upgrade only the winner. Retakes eat most of a budget - roughly seven of ten credits die on a bad take - so previs the complex shot as a white model first (7-10 retries down to 2-3), shoot 3-5s pieces and stitch, and change one variable per retry.
- Gate the first frame before you scale: wrong face or product, garbled text, malformed hands, CGI look. A failed render gets diagnosed in concrete words first - material, light path, aerial perspective, overlapping structure, a negative that did not hold - then repaired with those words. A blind re-roll with a new seed doubles the bill; only after the repair fails do you change the seed.
- Read the reference post before you write: pull its transcript and overview frame sheet locally so line timing and cut rhythm come from the real post, not a guess.
SHOT BOARD
1. [0.0-2.0] -, hard cut, Hook: A pattern breaks in the first two seconds. Start mid-action.
Prompt: Edit instruction: open on the strongest frame, no fade, no title. Cut on the first beat.
Foreground: a soft foreground shape thrown out of focus across the bottom third, out of focus, between the lens and the subject
Audio: Music starts on the cut.
On screen: [hook, 5 words]
2. [2.0-4.5] -, beat cut, Promise: The promise is stated in one line over the second-strongest frame.
Prompt: Edit instruction: one clip, one line, cut on beat two.
Foreground: a blurred doorframe edge across the left third, out of focus, between the lens and the subject
Audio: Music bed under voice.
On screen: [promise]
Line: [the promise]
3. [4.5-7.0] -, beat cut, Build one: Establish the pattern: same framing, same rhythm, third clip.
Prompt: Edit instruction: repeat the framing of shot 02 so the break later reads.
Foreground: the out-of-focus edge of a near wall across the right third, out of focus, between the lens and the subject
Audio: Bed continues.
4. [7.0-9.5] -, beat cut, Build two: The pattern holds; one detail changes.
Prompt: Edit instruction: same shot size, new subject detail, cut on beat four.
Foreground: a blurred doorframe edge across the left third, out of focus, between the lens and the subject
Audio: One new element enters.
5. [9.5-12.0] -, speed ramp, Proof: A visual proof: the thing working, in one unbroken shot.
Prompt: Edit instruction: speed ramp into the proof, hold the result for 12 frames.
Foreground: the out-of-focus shoulder of a chair back across the right third, out of focus, between the lens and the subject
Audio: Effect sound on the result.
On screen: [the proof]
Line: [what it proves]
6. [12.0-14.5] -, match cut, Escalate one: A shape match carries the eye into a new location.
Prompt: Edit instruction: match the shape, not the subject. Cut on the match.
Foreground: a blurred strand of hair across the left third, out of focus, between the lens and the subject
Audio: Music lifts.
7. [14.5-17.0] -, whip, Escalate two: Shorter shots, faster cuts, one new element per cut.
Prompt: Edit instruction: three clips in 2.5 s, each on a beat subdivision.
Foreground: a blurred doorframe edge across the left third, out of focus, between the lens and the subject
Audio: Percussion enters.
8. [17.0-19.5] -, hard cut, Turn: The obvious answer is wrong. Withhold the reveal.
Prompt: Edit instruction: cut to black for 4 frames, then the wrong answer.
Foreground: a hanging fabric edge, blurred across the top third, out of focus, between the lens and the subject
Audio: Music drops out for 4 frames.
Line: [the wrong answer]
9. [19.5-22.0] -, beat cut, Reveal: The real answer lands on the beat.
Prompt: Edit instruction: reveal on the downbeat, no transition.
Foreground: a blurred doorframe edge across the left third, out of focus, between the lens and the subject
Audio: Drop hits.
On screen: [the reveal]
Line: [the real answer]
10. [22.0-24.5] -, slow motion, Payoff: The hero moment, held long enough to feel earned.
Prompt: Edit instruction: slow to 40%, hold 1.5 s, then back to speed.
Foreground: a hanging fabric edge, blurred across the top third, out of focus, between the lens and the subject
Audio: Music peaks.
11. [24.5-27.0] -, hard cut, Button: The line the audience repeats.
Prompt: Edit instruction: one clip, one line, cut on the last beat.
Foreground: a soft foreground shape thrown out of focus across the bottom third, out of focus, between the lens and the subject
Audio: Music button.
On screen: [the button line]
Line: [the button line]
12. [27.0-30.0] -, loop cut, Loop: The last frame matches the first so the replay is seamless.
Prompt: Edit instruction: match the first frame within 2 px; no fade.
Foreground: a hanging fabric edge, blurred across the top third, out of focus, between the lens and the subject
Audio: Tail that loops into the first beat.
MOODBOARD PROMPTS
- Energy map (16:9): The reference energy curve drawn over its timeline, with the drop marked. This is the only thing the rebuild copies.
- Drop marker (16:9): The frame at the drop, with the hero shot that will land on it. One image, one decision.
- Cut strip (16:9): Twelve thumbnails in order, one per shot, with the beat number under each.
- Caption kit (9:16): Two caption styles tested at 390 px, one word-level pop and one clean lower third.
- Ratio check (1:1): The hero shot at 9:16, 1:1 and 16:9, so the re-frame is decided before the export.
- Beat grid (16:9): The waveform with downbeats, bar numbers and the drop marked. Every cut decision is made against this image, not by ear alone.
- Pattern pair (9:16): Two frames that establish the pattern the cut will break: same size, same position, one detail changed. The break only reads if the pattern is visible first.
- Drop frame (9:16): The hero frame that will land on the drop, at final crop, checked for subject position inside the platform UI.
- Ratio trio (16:9): The same hero frame at 9:16, 1:1 and 16:9, so the re-frame is decided once and reused for every export.
- Hook frames (9:16): Three candidate first frames, judged at thumbnail size. The winner is chosen before the rest of the cut is built.
- Silence marker (16:9): The four frames before the reveal marked on the waveform, so the silence is placed rather than discovered in the mix.
- Energy curve (16:9): The rebuilt cut's energy drawn over the reference's, so a flat stretch is visible before it is felt.
- Caption block (9:16): The caption block over the busiest frame, at 390 px, clear of the mouth and the platform UI.
- Cut log (16:9): Every cut on one line with its timecode, its beat and its reason. A cut with an empty reason column is deleted before the export.
- Loop pair (9:16): The first and last frame overlaid, with the offset measured in pixels. The loop is a measurement, not an impression.
SOUND AND MUSIC
The music is the clock. Mark the BPM and the beat grid before the first cut, put the drop on a marker, and let every cut either land on a beat or break one on purpose. Keep voice above the bed, duck 10 dB under speech, and leave four frames of silence before the reveal. Export with the music at -14 LUFS and true peak -1 dB. Treat the music as the timing authority: mark the BPM, the downbeats and the drop, then place picture against those marks. Duck the bed 10 dB under speech, leave four frames of silence before the reveal, and put the loudest moment of the mix on the drop. If a cut has no beat and no reason, delete it.
QC GATE
- Rights: The right to study the reference is recorded and the footage is yours; on fail: No right: stop
- Structure: The blueprint copies structure only, not look, subject or performance; on fail: A copied frame: rebuild it
- Beats: Every cut lands on a beat or a deliberate break; on fail: A random cut: re-time
- Drop: The hero shot lands on the drop within 2 frames; on fail: A late drop: shift the clip
- Hook: The first two seconds break a pattern; on fail: A slow open: re-cut the first shot
- Captions: Word-level, legible at 390 px, clear of the mouth; on fail: Overlap: move the block
- Grade: One grade for the whole cut; on fail: A shot out of family: re-grade
- Audio: Voice above bed, no clipping, silence before the reveal; on fail: A muddy mix: rebuild
- Ratios: 9:16, 1:1 and 16:9 are re-framed, not cropped; on fail: A lost subject: re-frame
- Disclosure: Reference credited and AI label where synthetic; on fail: Missing: block delivery
- Transcript: Captions are corrected against the script, not the auto-transcript; on fail: A wrong word: rebuild the caption layer
- Preset: Every export uses the saved preset for its ratio; on fail: A one-off setting: re-export
- Retention: The 3 s, 10 s and drop reads are recorded per platform; on fail: No read: the cut is not measured
- Pattern: The pattern is established before it is broken; on fail: A break with no pattern: add the establishing shots
- Silence: Four frames of silence sit before the reveal; on fail: A busy reveal: cut the bed for four frames
- Loop offset: The last frame matches the first within two pixels; on fail: A visible jump: re-cut the loop frame
FAILURE ATLAS
- Viewers drop at 3 seconds (The hook explains instead of breaking a pattern) -> Open mid-action and cut the first line
- Cut feels busy (Too many transitions, no pattern) -> Establish one pattern and break it once
- Reveal does not land (The drop was not marked before the edit) -> Mark the drop and re-time the last four shots
- Rebuild feels like a copy (Look and subject were copied with the structure) -> Replace the footage and the grade, keep the beat map
- Captions hurt retention (They cover the subject or lag) -> Move and re-time, word-level
- Export looks soft (Footage was upscaled in the edit) -> Match the canvas to the source
- Audio clips (Bed and voice summed without headroom) -> Re-mix with a limiter
- Replay does not happen (The last frame does not match the first) -> Match the first frame within 2 px and cut the fade
- Captions carry errors (Auto-captions shipped unedited) -> Correct against the script and re-export
- Exports look different per platform (Each export used its own settings) -> Apply the saved preset per ratio
- Cut feels random (No pattern was established before the break) -> Show the pattern twice before breaking it
- The rebuild copies the reference (Its grade, subject or performance travelled with the structure) -> Replace all three and keep only the beat map
STEPS
1. Confirm the right to study the reference and record it.
2. Analyze the reference: cut times, shot lengths, BPM, energy curve, drop timestamp, caption style.
3. Write the blueprint as structure only: hook, promise, build, proof, escalate, turn, drop, button.
4. State plainly what is not copied: look, subject, performance, script.
5. Log your own footage with duration, resolution and content.
6. Write the cut plan over the transcript: trim, split, concat, reorder, with a reason per cut.
7. Mark the beat grid and the drop on the timeline before placing clips.
8. Place one clip per shot at the same start and duration as the blueprint.
9. Reopen the cut at 10%, 35%, 65% and 90% and reject boring or repetitive sequences.
10. Land the hero shot on the drop and check it within two frames.
11. Add word captions and one grade; check at 390 px.
12. Export 9:16, 1:1 and 16:9 and write the retention notes.
13. Mark the BPM, downbeats and drop before placing a single clip, and keep the grid visible for the whole edit.
14. Correct the transcript by hand against the script before writing the cut plan.
15. Keep a cut log with a reason per cut; a cut with no beat and no reason is deleted.
16. Export every ratio from the saved preset and record the retention read at 3 s, 10 s and the drop.
17. Test three first frames at thumbnail size before committing to the hook.
18. Draw the rebuilt energy curve over the reference's before the final export and fix any flat stretch.
19. Measure the loop offset in pixels before export rather than eyeballing the join.
RESEARCH HARNESS
You are a production agent, not a prompt dispenser. Run this loop before you write a single shot:
1. Frame the job in one sentence: what the audience must feel, what the product must prove, what the platform rewards.
2. Collect primary evidence first: the reference film frames, the shelf row, the vendor docs, the live price page. Record the URL and the date. A blog post is corroboration, not proof.
3. Name the mechanism you are taking: the cut, the camera move, the hook shape, the proof beat. Take the structure, never the creator.
4. Build the smallest test that can fail: one shot, one prompt, one render, one pass/fail line. Do not board twelve shots before shot one survives.
5. Compare the render to the reference side by side. Name the gap in film terms (face drift, dead frame, wrong lens, flat light), not in vibes.
6. Change one variable at a time. Keep a run log: prompt, seed, model, duration, verdict.
7. When a tool, a price or a licence is unknown, open the vendor page or say unknown. Never fill the gap from memory.
8. Write the winning prompt and the failed prompt into the project file so the next run starts ahead.
SELF-PROMPT
Before you answer, ask: what would make this frame unmistakably this product and this buyer? After you answer, ask: which frame would a stranger screenshot? If neither answer is concrete, the work is not done.
OPENCODE SKILL STACK - Retention Edit
These are the installed skills this run loads. Load them in this order and do not add a skill that is not on the list.
- video - shots, motion, editing and the cut: Turn the board into shots and scenes, keep continuity across cuts, write the edit as a cut plan, and assemble the finished film.
- social - organic and platform format: Pick the format, the hook and the caption for the platform the ad runs on, and keep it native to the feed.
- copy-editing - copy and caption revision: Review the draft copy and captions, remove the AI tells and sharpen every line before it goes back into the next render.
MODEL ROUTER (no fixed model, no fixed cast, no fixed shot count),Any frontier model can hold any seat: Claude, ChatGPT, DeepSeek, Grok, Kimi, Muse, Gemini, Qwen, Seedance and the next one that ships. Use the model the buyer already pays for. The board, style.md, cast locks and receipts are portable between them.,The run is not a fixed number of shots or characters. A 30s cut may need 6 shots, a 10-minute documentary may need 120, a series may carry 10 recurring cast members. The core board is a starting point, not a cap. Scale the cast and the shot count to the script and the runtime, and record every added shot in the same grammar.,On a rate limit, outage or spend cap: save the run state (board, style.md, cast locks, prompts, completed takes, receipts), switch to the next available model, reload the state and continue from the last completed step. Never restart the run and never re-charge a completed take.,Language models own structure, copy, research and QC. Video models own motion and frames. Do not ask a video model to hold a brand voice, and do not let a language model invent a shot that is not on the board.
RUN ORDER
- INGEST: pull the ad from the source the buyer names. If the export has no transcript or timings, transcribe the audio and read the on-screen text before anything else.
- DECODE: write style.md. Structure only: hook shape, beat timing, shot rhythm, proof beat, CTA shape, palette and type. State what is copied and what is rebuilt.
- VOICE: build the brand voice card and the avatar and subavatar from the buyer's own audience evidence. Never reuse the source brand's voice.
- PLAN: write the board, the cut plan and the variant matrix. Price every take before it is generated.
- GENERATE: keyframes first, gate the first frame, then motion. Keep the cast, product and style locked between renders.
- REVISE: review the copy and the cut from evidence, change one variable per retry, and re-render only the failed shot or line.
- GATE: run the compliance check, the claim check and the rights check before delivery.
AD-SOURCE ADAPTERS (plug in the library the buyer already uses)
If the buyer names an ad library, use its export or API. If not, ask for one export, URL, screenshot, transcript or CSV. Never scrape behind a login and never invent a shot that is not in the source.
- Foreplay (https://www.foreplay.co): swipe file, ad copy, video downloads and timestamps. Plug: board export or API. Fields: ad URL, brand, copy, media, timestamps, notes.
- Meta Ad Library (https://www.facebook.com/ads/library): live ads, copy, start dates and platforms. Plug: ad URL plus a saved screenshot, transcript and caption. Fields: ad URL, page, platform, start date, copy, creative.
- TikTok Creative Center (https://ads.tiktok.com/business/creativecenter): top ads, hooks, music and formats. Plug: URL plus a saved video, transcript and caption. Fields: ad URL, brand, hook, video, music, format, engagement.
- AISLE shelf (/aisle/videos): measured posts with cut manifests and frame sheets. Plug: row URL or row id. Fields: row id, views, format, cuts, transcript, product.
- Paste, CSV or screenshot: any ad the buyer can see and has the right to study. Plug: ad URL, screenshots, transcript, caption, CSV. Fields: whatever exists; the adapter labels what is missing.
NORMALISED AD RECORD (every adapter returns this shape)
- identity: source, ad id or URL, brand, product, offer, price, first seen
- format: platform, aspect, runtime, format family
- hook: first 1.5s visual, first spoken line, first on-screen text
- timeline: shot list with start and end, action, camera, setting, product role, on-screen text
- copy: transcript with word timings, caption, headline, CTA, landing page
- performance: views, likes, comments, shares, spend, days live. Label measured, reported or missing.
- voice: tone, pace, point of view, lexicon, taboo words
- avatar: who the ad addresses, awareness level, beliefs
- subavatar: one named segment inside the avatar with 3am pain, enemy, scroll surface, proof and objection
- claims: every claim with its evidence, or flagged unsupported
- rights: usage and consent flags for footage, face, voice and music
LABEL WHAT IS MISSING
If a field is absent, write N/A. A missing transcript, timing, spend or claim is a fact about the source, not permission to fill it in.
OUTPUT CONTRACT
- Right-cleared reference
- Cut and beat analysis
- Declarative cut plan with gate marks and structured recovery hints
- Energy map and drop marker
- Rebuild with your own footage
- Word-level captions
- 9:16, 1:1, 16:9 exports
CHECKS
- The right to study the reference is recorded
- The blueprint copies structure only
- The beat grid and drop are marked before the edit
- Every cut has a reason
- The hero lands on the drop within two frames
- Captions are word-level and legible at 390 px
- All ratios are re-framed, not cropped
- The reference is credited
- The beat grid and drop are marked before the first clip is placed
- The transcript was corrected by hand
- Every cut has a logged reason
- Retention is recorded per platform, not inferred
- The pattern is visible before the break
- No flat stretch runs longer than four seconds
- The loop offset is measured, not eyeballed
REFUSE OR STOP WHEN
- No reference footage in the timeline
- No copied look, subject or performance
- No cut without a beat or a deliberate break
- No export without captions and credit
- No claim about retention that was not measured
Render budget, today
30s on the board · Seedance 2.5 720p
12 shots, 30 seconds of finished screen time, priced at today's published Seedance 2.5 720p rates. That is the render bill only: no retries, stills, music, voice or edit. The rate board carries every seller, the date each price was read and the calculator for a different length or resolution.
The stack
what each role uses
| Role | Tool |
|---|---|
| Cut planning | Claude |
| Edit | CapCut |
| Transcript edit | Descript |
| Dead-air removal | auto-editor |
| Captions | Submagic |
| Cutdowns | OpusClip |
Other packs
8 more
AI Drama Series
A recurring cast, an episode arc, and the same face in every shot.
10 shots / 60s episode · cast to scriptSong VSL · 9:16 · 12-stage journeyDrama Song Ad
A drama-style sung VSL, from script and song to timing map and locked cast.
8 shots / 30s song · cast to scriptCartoon · 9:16 · 3-7s fragmentsAI Cartoon Short
A cartoon with model sheets that survive every cut.
8 shots / 40s short · cast to scriptUGC · 9:16 · 100 ads a dayUGC Ad Factory
One shoot day, a hundred ads, each written for a real buyer.
6 shots / 30s ad · cast to script