---
name: aisle-clone-ad
description: Deconstruct any ad into a style spec, then rebuild the structure with your product, your cast and your proof - self-prompted, gated and ready to render. Use when you have a reference ad and want its structure without its words, face, brand or claims.
---

# Clone Any Ad

Clone the mechanism of an ad, not its surface. Strip the reference down to a four-part dramatic equation, measure its cut grammar, write a style.md that another agent can follow, then rebuild the same structure with a product, identity and proof you have the right to use. The clone passes only when a stranger sees the same structure and a different brand. Method adapted from mechanism-level imitation and the Higgsfield skills architecture; no source text is copied.

## Method

THREE-LAYER EXTRACTION
1. Strip the shell. List the reference's banned shell in banned-words.md: exact lines, brand words, the creator's face and voice, signature colours, slogans and any claim you cannot substantiate. These words never appear in the rebuild.
2. Keep the core. Write the four-dimensional equation: hook tension (what makes the first frame unavoidable), escalation mechanism (what raises the cost each beat), proof type (what kind of evidence closes the loop), and release (what the viewer is asked to do). Add the measured cut grammar: shot lengths, cut rate, first-frame hold, text-on-screen timing and audio role.
3. Reskin across domains. Same equation, new product, new cast, new proof, new voice. Four collision questions before generation: would the original creator recognise a line? does any banned word survive? is the product role the same while the object is different? would a stranger name the same structure but a different brand?

HIGGSFIELD MODEL ROUTING (adapted)
When the target model exposes ad_reference_id, hook_id or setting_id, pass the ids recorded in style.md. When it does not, put the same fields in the prompt text. Route live: cinematic_studio_3_0 primary with veo3_1 fallback for film-grade looks, marketing_studio_video with seedance_2_0 for shopping and marketing, wan2_6 with seedance_2_0 for panel or storyboard sources, kling3_0 with grok_video for social hooks. Re-verify every parameter against the live schema before spending.
Platform durations are model properties: veo3_1 uses 4, 6 or 8 second clips; wan2_6 uses 5, 10 or 15; most others are 4-15. A longer ad is a chain of confirmed clips, never one imaginary 60-second call.

## Self-prompt loop

SELF-PROMPT LOOP (run before every generation, and after every result)
1. Restate the job as a stranger would send it: product, proof, buyer, runtime, ratio, cap. A missing field is a stop, not a guess.
2. Ask what a skeptical stranger would need to SEE in the first frame to believe the promise. Answer with a drawable object and a number, never an adjective.
3. Write three candidate prompts for the next step. Score each against the locked contract, the sheets and the gate that step serves. Keep the one with the most verifiable detail; log the two losers with the reason they lost.
4. After the result, write the prompt you would hand a fresh agent to reproduce it exactly. If that prompt drops a lock, a gate, a number or a reference job, the step is not finished.
5. Change one variable per retry. A retry that changes two variables teaches nothing and is logged as a failed experiment.


## Generation and chain contracts

GENERATION CONTRACT (from the Higgsfield canonical opt-in flow)
- Resolve the model's live schema before assembling parameters; a cached catalog is a lead, not the source of truth. Re-verify if a parameter is rejected.
- Assemble prompt, aspect ratio, duration and model-specific fields, then SHOW the assembled prompt for revision. The user approves prompt text, not just a final render.
- If the job needs a reference image: media_upload returns pending_id, media_confirm returns confirmed_id, and only the confirmed_id goes into input_files with a declared role (image, start_image, end_image, video, audio). Never pass a pending_id.
- Show balance and credit cost, then wait for an explicit YES before any paid call. No YES means no generation; offer a revision and return to prompt assembly.
- Generate once, capture the job_id, then poll status with backoff. Never re-submit generation to find out whether a job finished: that buys a second job.
- A failed job reports its reason and original job_id; a retry goes through the confirmation gate again.
- Duration is a property of the model. Split longer cuts into chained clips and confirm each clip separately.

CHAIN CONTRACT (for cuts longer than one model clip)
- Map the whole runtime to clips before the first generation. Each clip has a start, an end, an emotion, a product role and an exit frame.
- Generate and gate the exit frame of clip N; that approved frame becomes the entry frame of clip N+1. A chain with an ungated handoff frame is a chain of unrelated clips.
- Hold face, wardrobe, product, palette and lens in the sheets. Never re-describe a character from scratch in a later clip.
- Keep a runlog row per clip: prompt, model, seed, reference ids, duration, verdict, cost, and the gate that approved it.
- Vary within a section without changing the locks. The same prompt twice is a wasted generation, not a safety measure.

## Three-engine pipeline

STORY BRAIN - emotional skeleton and the four-block storyboard (engine 1 of 3)
Runs before any image or video call. It produces the paper the other two engines consume verbatim. Input: one beat, the product truth, the buyer, and for the drama run the Story Spine. Output: four labeled blocks in this order, carrying exactly one emotional beat.
1. CHARACTER SHEET - silhouette, palette, wardrobe, signature prop, posture default, eye behavior. Copied byte-for-byte from the Character Lock when one exists.
2. SCENE PROMPT - one emotional beat, the want, the obstacle, a specific stylistic anchor (a named film + year, never "Pixar-style"), and how the environment reacts to the character.
3. IMAGE PROMPT - one still at the peak of the beat: composition, eye line, palette. No motion verbs.
4. VIDEO PROMPT - anticipation Xs / action Ys / reaction Zs with explicit beat counts, plus the secondary motion (hair, cloth, dust) that reacts on each beat.
Self-review, five gates: exactly one emotion; the anchor is a film + year; the video block names beat counts; the image block names an eye line; the character sheet matches the lock verbatim. Any no revises that block before the handoff. Refuse compound moods ("sad but hopeful and tired"), generic style words, neutral faces with no feeling or eye line, environment as backdrop, collapsed anticipation and action, and paraphrased identity tokens.

CHARACTER / CONSISTENCY LAYER - identity lock before generation (engine 2 of 3)
Mandatory before the second scene drafts, and again on any drift. A pose, framing or lighting change moves most pixels while identity holds, so pixel similarity is the wrong test. The control is verbatim tokens plus a human side-by-side against a reference frame.
Lock file characters/<id>.json, every field mandatory: id, name, subject_class (humanoid | creature | vehicle | abstract | object), silhouette (one-line read from 30m), palette (2-5 hex), wardrobe, signature_prop, posture_default, eye_behavior, face (age band, skin, hair, marks), voice_note (explicit null when not applicable), reference_frame, version.
Downstream prompts cite this file and copy its tokens verbatim, never paraphrase. Tokens are immutable until a revision bumps version and adds a CHANGELOG line (v<n> · YYYY-MM-DD · reason · scenes-to-rerun). Every scene that used the old version is re-reviewed by eye after a bump. Always set subject_class for non-humanoid subjects or the lock reads as humanoid downstream.
After the first successful render, copy the best full-face full-body frame to characters/<id>.ref.png and point reference_frame at it. Bootstrap scenes have no reference frame; the JSON is the only lock until then.
Recurring hero products, creatures, vehicles and signature objects get the same lock with subject_class set to object, creature or vehicle. A real person's identity or voice requires a cited likeness release before the JSON is emitted; recognised public figures carry the harder gate. Refuse and surface the path rather than silently sanitising it.

VISUAL ENGINE - storyboard block plus character sheet to assets, model chosen automatically (engine 3 of 3)
Consumes the four-block storyboard and the lock, resolves the model live, and generates gated assets. It never invents style; it chooses the model that best serves the locked block.
Route: cinematic_studio_3_0 primary with veo3_1 fallback for a film-grade frame; marketing_studio_video with seedance_2_0 for shopping and marketing; wan2_6 with seedance_2_0 for panels, boards and comics; kling3_0 with grok_video for social hooks; nano_banana, gpt_image or flux_kontext for an image plate; the higgsfield 3D-CGI route (or Maxfusion) for stylised volume once the character sheet is locked. When a provider exposes ad_reference_id, hook_id or setting_id, pass the ids recorded in style.md; otherwise put those fields in the prompt text.
Before every paid call: verify the live schema, assemble prompt plus ratio and duration, show the prompt for revision, surface balance and cost, and wait for an explicit YES. Upload reference media as an atomic media_upload then media_confirm pair; only confirmed ids enter input_files with a declared role. Generate once, capture the job_id, poll status with backoff, and never re-submit generation to check a job.
Duration belongs to the model: veo3_1 clips are 4/6/8s, wan2_6 are 5/10/15s, most others 4-15s. Longer ads are chained clips. Carry the approved exit frame of clip N into clip N+1, keep the lock tokens in every prompt, and gate each first frame before motion. A 45-second drama is eight chained clips; a 90-second drama is fourteen.

## Worked showcase

WORKED SHOWCASE - the three engines on one ad clone
Reference: a 12-second fragrance carousel. Rebuild for Arc Desk Lamp, $39, with a dated receipt and a 01:30 assembly timer as proof; 9:16; cap 40.
ENGINE 1 OUTPUT - four blocks:
- CHARACTER SHEET: copied verbatim from the lock below, no paraphrase.
- SCENE PROMPT: relief at 23:10 when the desk corner becomes readable; want = finish the work; obstacle = overhead light washes the page; anchor = "Up (2009)" warm practical, never "Pixar-style"; the environment reacts as dust becomes visible inside the beam.
- IMAGE PROMPT: 50mm, eye line to the switch at frame centre, 3200K practical, one blurred coffee cup in the near third; no motion verbs.
- VIDEO PROMPT: anticipation 2s (thumb hovers, lamp dark) · action 3s (switch clicks, beam lands) · reaction 2s (shoulders drop, dust drifts); secondary motion: sleeve settles on beat 2, dust on beat 3.
ENGINE 2 OUTPUT - identity lock:
{
  "id": "mara-arc-lamp",
  "name": "Mara",
  "subject_class": "humanoid",
  "silhouette": "compact upright read, shoulders slightly forward at a desk",
  "palette": ["#2b2f36", "#c9a15b", "#e8e2d6"],
  "wardrobe": "grey work shirt, rolled sleeves, no branding",
  "signature_prop": "short dark hair with a scar on the left thumb",
  "posture_default": "leaning in, forearms on the desk",
  "eye_behavior": "one slow blink, glance to the switch before acting",
  "face": "30s, medium skin, short dark hair, scar on left thumb",
  "voice_note": null,
  "reference_frame": null,
  "version": 1
}
ENGINE 3 ROUTE: marketing_studio_video primary / seedance_2_0 fallback at 9:16; first clip 5s, exit frame gated, second clip 10s carries the same lock tokens; explicit YES before each call; job_ids polled, never resubmitted. A first-frame failure (fused thumb) is diagnosed as hand geometry, one negative changed, retry logged.
FALSIFIER: if a stranger names the fragrance brand instead of the desk lamp, the clone failed even though all three engines returned clean artifacts.

## Inputs

- `reference` _url_: Reference ad: post URL, AISLE row, video file, screenshots or transcript. Example: `https://sauce-ops.appskins-studio.workers.dev/aisle/videos`
- `product` _text_: Your real product, price, what it does and what it does not do
- `cast` _text_: Your presenter identity or a request to generate an original one
- `proof` (optional) _text_: The artifact that survives a screenshot, with its source and date
- `runtime` (optional) _number_: Target runtime in seconds. Example: `15`
- `ratio` (optional) _text_: Primary ratio. Example: `9:16`
- `model` (optional) _text_: Target video model, or blank to route live. Example: `veo3_1`
- `cap` (optional) _number_: Hard spend cap enforced before each confirmed generation. Example: `60`
- `destination` (optional) _text_: Destination surface: Reels, TikTok, Meta, YouTube Shorts or pre-roll. Example: `Reels`

## Live stack

Call stack_get({ id: "recreate" }) for the full record with every option for each seat.
- Direction and motion: Higgsfield | $0: Wan 2.2 | also does this job: Sora, Google Veo, Kling, Luma Dream Machine
- Voice: ElevenLabs | $0: Kokoro-82M | also does this job: Resemble AI, Cartesia, Deepgram, LMNT
- Lipsync and avatar: HeyGen | $0: LivePortrait | also does this job: D-ID, Hedra, Argil, Captions
- Edit and captions: Submagic | $0: auto-editor | also does this job: CapCut, Adobe Premiere, DaVinci Resolve, Kapwing
- Music: $0: MusicGen | also does this job: Suno, Udio, Soundraw, Beatoven
- Product and pack shot: Zendrop | $0: ComfyUI | also does this job: Sora, Google Veo, Kling, Luma Dream Machine
- Product sourcing: AutoDS | $0: Alibaba | also does this job: CJdropshipping, Faire, 1688, Made-in-China

## Run

1. **Anchor the reference** - `resolve_input({ input: "{{reference}}" })`
   Resolve the reference to a product, niche and language. If it is an AISLE row, record the row id and date. If it is a bare URL, say what could and could not be verified.
   Output: source.md with reference id, URL, date and resolution status
2. **Pull the measured cuts** - `reverse_video({ row_id: ROW_ID })`
   Read cut timings and frames when the row supports it. A pending manifest is a real answer: label it pending and fall back to a written beat estimate instead of inventing timestamps.
   Output: measured shot table or a clearly labelled inferred table
3. **Get the paste-ready brief** - `recreation_brief({ row_id: ROW_ID })`
   Take beats, camera, edit rhythm and prompt blocks. Keep measured fields separate from inferred fields in every downstream document.
   Output: brief with measured and inferred columns
4. **Write the banned shell** - `SHELL STRIP`
   List exact lines, brand words, signature palette, creator face and voice, and unsupported claims from the reference. Everything on this list is forbidden in the rebuild. Permission changes nothing about claim support.
   Output: banned-words.md with a reason per item
5. **Extract the dramatic equation** - `MECHANISM EXTRACTION`
   Write hook tension, escalation mechanism, proof type and release in four lines. Add the measured cut grammar: cut rate, average shot length, first-frame hold, text timing, audio role. This equation, not the caption, is what gets cloned.
   Output: mechanism.md with the four dimensions and the cut grammar
6. **Write style.md** - `STYLE SPEC`
   One paste-ready spec another agent can follow: logline, six-second hook rule, palette, lens, lighting, grade, edit rhythm, typography, sound map, duration map, model route and the four collision questions. No field may say "cinematic" without naming light, lens and stock.
   Output: style.md with every slot filled
7. **Place your product and cast** - `product_placement({ product: "{{product}}", scene: "FIRST SHOT", aspect: "{{ratio}}" })`
   Generate hero, in-use and detail placements for the real product. Lock the presenter identity and reject any plate that hides the label or changes the product shape or colourway.
   Output: placement coordinates, prompt_fragment, negative_prompt, cast lock
8. **Route the model live** - `models_explore`
   Resolve the live schema for {{model}} or browse models when it is blank. Capture accepted aspect ratios, durations, reference limits and ad_reference_id / hook_id / setting_id support. A cached catalog is a lead, never the source of truth.
   Output: model-route.md with the verified schema and the fallback
9. **Self-prompt three candidates** - `SELF-PROMPT`
   Write three candidate prompts for the first frame, score them against the equation, the banned shell and the cast lock, keep the strongest and log the two losers with their reasons.
   Output: prompt-log.json with three candidates, scores and the chosen prompt
10. **Confirm before you spend** - `COST GATE`
   Show balance, credit cost, assembled prompt, parameters and reference ids. Wait for an explicit YES. If the user says no, offer a revision and return to prompt assembly. No YES means no generate call.
   Output: confirmation record with the exact prompt approved and the cap
11. **Attach references atomically** - `media_upload then media_confirm`
   When a reference image or video is required, upload to pending_id, confirm to confirmed_id, and attach only the confirmed id with a declared role. Never pass a pending_id to input_files.
   Output: confirmed reference ids with roles
12. **Generate and gate the first frame** - `generate with the chosen model`
   Generate once with the approved parameters, capture the job_id, poll status with backoff, and gate the first frame against the hook, the product and the cast. A failed frame is diagnosed in concrete words and logged before a single-variable retry.
   Output: keyframes/ with the gated first frame and runlog.json
13. **Chain the remaining clips** - `CHAIN CONTRACT`
   Generate clip N+1 from the approved exit frame of clip N. Re-check face, wardrobe, product and palette at every handoff. Never re-submit a generation to check a job; poll with status.
   Output: gated clip chain with a handoff frame per join
14. **Edit, caption and recompose** - `AGENTIC EDIT`
   Cut on the hook, retime to the exact runtime, burn text in the edit layer, caption per platform, recompose 9:16 into 4:5 and 1:1 by reframing rather than regenerating, and export FCPXML, OpenTimelineIO or a CapCut-ready folder.
   Output: edit/ with captions, ratios, sound map and exports
15. **Run the clone-fidelity gate** - `MACHINE CHECK C1-C16`
   Every line is pass/fail and a failure blocks delivery. Check source, shell ban, equation, measured cuts, model schema, explicit YES, poll discipline, handoff gates, product visibility, proof support, runtime, ratios and platform policy.
   Output: qc.md with C1-C16 and blocking failures named
16. **Package the handoff** - `PACKAGE`
   Put every contracted file in one folder with the run log and the QC result. A stranger must be able to rebuild the clone from style.md and the sheets alone.
   Output: delivery folder with all ten contracted files

## Worked example

### Worked example - fragrance carousel to a $39 desk lamp
- **Input**: reference: a real 12-second fragrance carousel picked from /aisle/videos; product: "Arc Desk Lamp, $39, three colour temperatures, flat-pack, two-year receipt"; cast: original presenter, no likeness of the reference creator; proof: the receipt and a timer showing the 90-second assembly; runtime 15; ratio 9:16; cap 40.
- **Contract**: banned-words.md: the fragrance brand, the word "lost cherry", the exact opening caption, the creator face and the moody voiceover. mechanism.md: hook tension = a dark desk and one unlit corner; escalation = three failed work sessions at 11pm; proof type = the timer and the receipt in one frame; release = one tap to the product page. style.md: warm 3200K practical, 50mm, slow push, hard cut on the click, no logo end card.
- **Beats**: 0.0-2.5 GRAB: finger hits the switch, corner stays dark, lamp off-screen. 2.5-6.5 FACE IT: three coffee cups and a clock at 23:10, one line: "Another night squinting at a desk." 6.5-11.5 PROOF: the lamp snaps on, the timer reads 01:30, the receipt lies beside it - evidence and conclusion in the same beat. 11.5-15 AFTERSHOCK: the desk is readable, one CTA line, end on the transformed desk.
- **Gate**: C1 row id recorded; C2 banned shell has six items; C3 zero banned words; C4 original identity; C8 product at 1.8s; C10 receipt dated; C11 veo3_1 schema verified; C12 explicit YES logged; C14 handoff frame gated; C16 policy passed. One failed first-frame render (finger fused with switch) was diagnosed as a hand-geometry fault, one variable changed, and the retry logged.
- **Falsifier**: If a viewer names the fragrance brand when shown the rebuild, the clone failed even though every machine check passed.

## Research harness

RESEARCH HARNESS
You are an operator, not a form. Run this loop before you act:
1. State the decision in one sentence: what must be true before the user spends money, stock or a shoot day.
2. Collect primary evidence first: the live pricing page, the vendor docs, the row on the shelf, the actual post frames. Record the URL and the date. A listicle is a lead, not proof.
3. Name the assumption that would break the answer. Test that one first.
4. Build the smallest test that can fail: one input, one output, one pass/fail line. Do not scale before the test passes.
5. Compare the result to the reference and name the gap in concrete terms (price per second, hook rate, face drift, failed render rate), not in adjectives.
6. Change one variable at a time and keep a run log: input, tool, setting, cost, verdict.
7. When a price, rate or licence is unknown, open the source or say unknown. Never fill the gap from memory.
8. Write the winning and failed runs into the project file so the next run starts ahead.

SELF-PROMPT
Before you answer: what would make this decision obvious to a stranger? After you answer: which number or frame proves it? If neither is concrete, keep digging.

## OpenCode skill stack



## Ad-source adapters

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.
- TrendTrack (https://www.trendtrack.io): winning ads, hooks, spend and engagement, creative library. Plug: CSV export, board share link or API key. Fields: ad URL, brand, hook, copy, media, timestamps, spend and engagement.
- Brandsearch (https://www.brandsearch.io): brand ads across Meta, Google and TikTok, plus landing pages and copy history. Plug: board export, URL or API. Fields: brand, ad URL, platform, copy, creative, landing page, first seen.
- GetHookd (https://www.gethookd.ai): winning ads, hooks, scripts and transcripts. Plug: collection export, CSV or API. Fields: ad URL, hook, script, transcript, creative, format, engagement.
- 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

- source.md - reference id or URL, date, what was measured and what was inferred
- banned-words.md - shell items with a reason each
- mechanism.md - four-dimensional equation plus the measured cut grammar
- style.md - paste-ready style spec with the model route and collision questions
- sheets.md - product sheet, cast lock and negative prompt
- board.md - timecoded shot table with product role, text and prompt
- keyframes/ - gated first frame and the approved exit frame per clip
- runlog.json - candidate prompts, model, seed, reference ids, verdict and cost
- edit/ - captions, recomposed ratios, sound map and exported timeline
- qc.md - C1-C16 with blocking failures named

## Files

- `source.md` - reference id, URL, date, measured versus inferred
- `banned-words.md` - shell item and the reason it is forbidden
- `mechanism.md` - hook tension, escalation, proof type, release, cut grammar
- `style.md` - logline, palette, lens, grade, edit, sound, typography, duration map, model route
- `sheets.md` - product sheet, cast lock, negative prompt
- `board.md` - shot, start_s, end_s, action, camera, text, prompt
- `keyframes/` - gated first frame and exit frame per clip
- `runlog.json` - prompt candidates, model, seed, refs, duration, verdict, cost
- `edit/` - captions, ratios, sound map, export paths
- `qc.md` - C1-C16 with blocking failures named

## Cost and time

The buyer sets a hard cap and it is checked before each confirmed generation. Route live, draft on the fast tier, gate the first frame, and upgrade only the winner. A re-submitted generation is a new charge and a failed experiment.

## How this run is judged

A clone that copies the caption has failed. The unit of cloning is the mechanism: what makes the first frame unavoidable, what raises the cost, what kind of proof closes the loop, and what the viewer is asked to do. Keep measurement and inference in separate columns so an inferred cut is never quoted as a fact. The banned shell is the copyright and consent boundary; the self-prompt loop is the quality boundary; the explicit YES is the spend boundary. All three have to hold before a frame is paid for.

## Stop condition

Stop at the output contract. If the banned shell, the model schema or the explicit YES is missing, stop before generation. A failed C-line blocks delivery.

## What makes this run fail

- **The clone is the reference with a new logo** - Only the caption was changed. Fix: Re-extract the four-dimensional equation and rebuild from it
- **A banned line survived** - The shell strip was written after the script. Fix: Write banned-words.md before any new copy and gate on it
- **The first frame is paid for and wrong** - The prompt was never shown for revision or no explicit YES. Fix: Show the prompt, then wait for YES before generation
- **The chain falls apart at clip three** - No approved exit frame was used as the next entry. Fix: Gate the handoff frame before generating the next clip
- **The job was charged twice** - Generation was re-submitted to check status. Fix: Poll with job_status and keep the original job_id

## Rules

- Evidence before prose. Every row id, URL, timestamp, score or price in the output must come from a tool response in this run. If it did not, label it as an assumption.
- Keep measurement and judgment apart. View counts, cut timings and scores are measured. "This will convert" is a judgment and must read like one.
- No invented links, codes, prices or stock levels. Use only the links the registry returns.
- No artificial cap on shots, scenes or AI-generated cast. Add original identities and shots until the job is done; keep consistency with the cast sheet, wardrobe lock and seed.
- Deduplicate to concepts. Reposts and re-edits of one ad are one data point, not five.
- Ask once for a missing required input. If a tool returns nothing, report the empty result and stop. Never fill a gap with a guess.
- Organic first. Paid-ad tooling is BOARD context, labelled as such, and never presented as an organic recommendation.
- Finish with the deliverable. Do not close with a summary of what you did.

## Shot prompt grammar

- Every shot prompt runs Subject, Action, Camera, Style, Constraints (S-A-C-S-C) in that order: 50-80 words, three sentences, one visual promise. "Cinematic" is not a quality switch - name the light, the lens and the stock.
- Every shot carries a foreground: one blurred object across about a third of the frame, between the lens and the subject, plus one atmospheric layer behind it (fog, dust, steam, a volumetric beam). Say blurred or the model renders it sharp.
- Length matches the file: second-level segments (0s-3s: ...) that add up to the requested duration, because a timeline that does not match makes the model drop the ending. Chain clips by generating the last frame of clip N as the first frame of clip N+1.
- One job per reference image, named in the prompt itself; a reference that already shows the action needs only "strictly follow clip 1". Cap references at 30 images, 10 videos and 10 audio.
- Draft at 480p or the vendor fast tier, gate the first frame, diagnose a failed render in concrete words (material, light path, structure, a negative that did not hold) before re-rolling, and change one variable per retry. Retakes are where the budget goes, not the render.
- Attach the cast sheet to every prompt and never re-describe the character from scratch. Play emotion as body detail, and keep sprints, jumps and rolls out of a single shot - hand the energy to the cut.

## Checks

- The reference id, URL and date are recorded; unmeasured fields say inferred.
- The banned shell is non-empty and no banned word survives in the new script.
- No creator face, voice, brand mark or unsupported claim was copied.
- The four-dimensional equation and the cut grammar are both present.
- Shot timings come from a tool response or are labelled inferred.
- The cut rate is within 15 percent of the reference or the deviation is justified.
- The first frame carries the hook visually and the product is visible by second 3.
- The proof artifact has a source and a date, or the promise is an invitation to try.
- The model schema was verified live before generation.
- ad_reference_id, hook_id and setting_id were passed when the model supports them.
- An explicit YES was recorded before every paid generation.
- Reference ids are confirmed ids with declared roles, never pending ids.
- Jobs were polled with status and never re-submitted to check completion.
- Every clip handoff used an approved exit frame and held the locks.
- Ratios were recomposed by reframing and captions live in the edit layer.
- Platform policy passed before spend.

## Refuse or stop when

- Ask once for a missing required input and stop; do not invent a product, a price, a face or a claim.
- If the request is to copy a creator's face, voice or identity, require consent or a licensed actor; otherwise generate an original identity or a faceless cut.
- If no proof exists, remove the claim or turn the promise into an invitation to try.
- If the live model schema rejects a parameter, re-verify it before retrying and log the change.
- If spend would exceed the cap, stop before the next confirmation and report the remaining budget.
- If the reference is a regulated ad (health, finance, income), carry no claim into the rebuild without substantiation.

## Next

emotion-ad - The style spec defines structure; the emotion run turns that structure into a gated 15-second cut with a single feeling and a proof artifact.

## Connect

MCP endpoint: `https://sauce-ops.appskins-studio.workers.dev/api/aisle?op=mcp`

Read the full record over MCP with `workflow_get({ id: "clone-ad" })`.
