Human version

Summary. Turn a researched advertising promise into a feasible video in the requested format. First load angle research, then bind it to the agreed story, visuals, timing and assets. Write and check the complete concept before the user reviews it. The output is one recommended script, its visual plan and up to two viable alternatives. A model score is a quality screen; ad results establish performance.

---

Agent version

When to use

Use before production for voiceover, dialogue, chat, cards or lyrics. For a silent format, still plan its visual promise and payoff; skip spoken copy, not creative strategy.

When the user supplied exact lines, keep them verbatim. Check them in report-only mode and raise material timing, claim or format conflicts. Rewrite only with their authorization. This skill adds no review round of its own; the script goes into the run's existing review.

Inputs and precedence

Read this skill's files reference before creating working files. Required context:

• The brand and exact product, current facts, voice and prohibited claims.
• Audience, campaign objective, offer and CTA, using saved context and brief first.
• The selected format/route and any actual template recipe: capabilities, choices, assets and inputs.
• A validated angle context from ad-angle-miner, including source records.
• Reference footage or its observed beat map, plus available brand assets.
• The current sourced creative brief and applicable project/brand decisions, including exact rejected wording, pronunciations and user-locked copy.

The user's explicit direction and chosen angle are fixed. Actual route capabilities determine what is possible. A supplied reference controls the agreed custom structure; templates supply compatible components. Expose concrete gaps and feasible alternatives within existing review, without silently replacing angle or required demonstration. A reference provides a storytelling structure; the product's facts provide claims.

Step 0. Load the angle research, explicitly

Load ad-angle-miner with catalog_fetch {type: "skill", slug: "ad-angle-miner"} unless its instructions and handoff reference are already loaded; fetching it does not run research. Reuse an existing angle bank (this run, the brief's pointer, the product's saved angle-bank.json, then legacy miner output), normalize it to video-angle-bank.v1 with the preparation script, and keep the selected angle. With no usable bank, run the miner in video mode for this audience, product and template. Follow angle research for the order, the checks and what to record; never claim the miner ran when it did not.

Step 1. Bind reference, format and production route before writing

Read the actual selected route/recipe, not just its catalogue card. For custom work, carry the agreed reference structure and real capability limits into shape.json:

• Actual selected template/route identity, format, story mechanism, audience and objective.
• Ordered slots, speakers, text kind, durations and explicit word/character/line limits.
• Product entrance, proof device, payoff and CTA placement allowed by this recipe.
• Available asset ids, permitted visual modes and any required demonstration.
• What can change and what stays fixed, including silence, music or lyric constraints.

For custom work, read reference and performance. Map actual source intervals to adapted lines, actions/proof and sound. Preserve agreed reference structure rather than force nearest-template scene count or pace. Keep speech separate from music/ambience/effects and intentional total silence.

For speech, read pacing. Set the desired cadence from the brief's explicit delivery direction or observed reference audio, then the recipe's default. Preserve explicit recipe limits and record any conflict. Without usable evidence, use a provisional 3 words per second for conversation or 2.5 for slow narration, labelled as a fallback.

Budget each spoken beat from its speech window, excluding pauses, silent demonstrations and end cards. Rates can differ across beats and speakers. Record the rate's source and keep recipe estimates separate from observed rendered delivery. Carry this plan into production direction. Read chat and cards against their own display time; lyric timing follows bars and syllables. A word estimate does not prove that a generated read fits. Do not change the recipe to express one run's target or promise forty words in fifteen seconds.

Examples of adaptation, not a fixed format catalogue:

Recipe	Script must do	Failure to reject
iMessage conversation	Build a believable exchange and reveal through short turns	A sales monologue split into bubbles
Product demonstration	Pair a supported benefit with a visible action or test	A testimonial over unrelated beauty shots
Creator narration	Match what is said to what appears at that moment	Invented personal results or unavailable footage
Silent product loop	Express one visual promise through reveal and final card	Forcing a spoken problem-solution script into it
Music or lyrics	Make the product story work within musical timing	Treating a lyric as ordinary voiceover

If the recipe lacks enough information, inspect the demo and capability instructions. Record any provisional timing. Resolve essential missing assets before approving a production plan that relies on them. In a requested writing-only test, specify the exact missing capture or reference and mark the draft not production-ready. A planned demonstration is not verified proof; never generate a fake UI to fill the gap.

Step 2. Form the creative brief

Before drafting, read current product evidence and the project's saved history, then save creative-brief.json (files reference): variant, buyer situation, supported mechanism, objective, offer (or "no offer supplied"), CTA, constraints, prior decisions with exact rejected wording, locked words and delivery intent. Pass it to prepare_angle_context.py --brief. Establish the message before choosing a reference. Buyer language sets voice; it never substantiates a claim, and fiction never poses as a real review. Follow the creative brief and buyer perspective and product explanation.

Step 3. Learn the reference's persuasion

Read actual reference/demo lines and timing, inspecting complete picture and sound. Frames/transcripts alone cannot establish delivery, continuous action or soundtrack; record missing duties and keep source observations distinct from adaptation. Add at most three relevant references from existing research. For each record the opening visual and words, beat order, product entrance, proof, objection, payoff and CTA. Record what transfers and what belongs to the source brand.

Match the reference's persuasive purpose to the creative brief, not just its category or visible action. Complete inspection establishes what is in the source; it does not make the source appropriate for this brand. Reject an unsuitable transfer before writing.

For dialogue, load dialogue-writing. A format description, product page or imagined beat map is not an observed conversation. Reuse the miner's sources first; retrieve missing references when access is available.

For a street ad, inspect a complete commercial clip, including picture and speech. Record opening, setup, engagement, hook, product connection and payoff; mark unseen approaches unknown. Match the actual offering and interaction. Keep timestamped words and actions, label paraphrases and retain the full transcript. Snippets, radio interviews and seed summaries cannot replace this observation. Run the selected renderer's street-script-writing selector (in the render-street-interview skill). Needs-reference or unsupported-route stops approval; report the gap and supported alternatives.

For a podcast, an appropriate publisher transcript can establish turn structure. Editorial conversation may be provisional evidence of that mechanic, with missing commercial context explicit. Neither a transcript nor sampled frames establishes vocal delivery; record what was actually heard or viewed. Authorship and ad performance remain unknown unless independently evidenced.

Keep performance labels honest: running duration and variants show advertiser persistence; views show reach or engagement. Neither proves conversions, profitability or causal lift. Prefer measured first-party results, with their context. Never call an untested reference a proven winner.

Step 4. Choose distinct angle–execution pairs

When the angle is open, work from up to 10 researched candidates. Filter unsupported promises and incompatible executions first, then write 4-5 concepts with different promises and openings, each a different angle type (angle_type, listed in the files reference) chosen by what the viewer already knows; see hook families. When the angle is selected, keep its id, promise and angle type, and write 4-5 executions that differ in hook family and opening. Batch assignments differ where the brief permits; use the bank once per brand and product.

Rank relevance, product credibility, template fit, feasible proof and hook payoff before freshness: a clear, supported familiar angle beats an obscure weak one. If the evidence supports fewer than four distinct concepts, write fewer and say why; never pad with paraphrases.

Step 5. Write words and visuals together

Read hook families before writing hooks, voice before any line (draft 3 hooks each), and scene and hook before writing dialogue. Plan the encounter, the literal first 0–3 seconds and the story's development/payoff first. Separate the participant's filming context from the context actually given to a new viewer. A complete question can be the hook; a reply-first opening needs its subject or task established in the retained picture, words or rendered text.

For every beat write the exact line, its on-screen words (13 or fewer; 8 in beat 1), speaker, visual action, production mode and existing asset ids when applicable. Beat 1's picture, line and on-screen words must say the same thing. Keep a proof plan and a claim ledger with current product fact ids. Every hook variant must lead into the same body's promise; repair the body when the promise changes.

Tie each material claim to the moment its demonstration is visible. Include action completion, recognition and text-reading time in the existing scene plan. Inspect the actual selected media before calling that plan feasible; metadata is a shortlist. Keep product truth/style references separate from usable production footage. If the needed proof is absent, narrow the claim or specify the missing capture; do not treat a simulated result as evidence. A silent format communicates through exact on-screen words and visible actions, with no invented speech.

• Make the first moment intelligible and worth watching, with or without sound.
• Keep one main promise and pay it off visibly.
• Use the product's actual mechanism, differentiator or demonstration where the recipe fits.
• Match dialogue, cards and lyrics to their medium and the brand's voice.
• Make the offer and CTA clear. Respect the recipe's ending and permitted CTA placement.
• Brand visibility depends on objective and format; early is not wrong by itself.
• Include only supported product claims. A source id is traceability, not a semantic check.

For conversational formats, write the connected exchange before allocating timed slots. Reject disconnected spoken headings or bullet stamps; short natural replies remain valid. Caption chunks and scene boundaries do not fracture the underlying thought. Ordinary UGC starts brisk and conversational; reference-supported calm proof/pauses remain valid. Use human speech and behavior, with matched language evidence and the human's rejected examples. Establish ordinary immediate motives, bounded speaker knowledge and why each question or recommendation happens now. Give each person an intention and a reason to respond to the previous turn. Then trim and add product inserts, captions and the CTA. Do not turn every slot into a sentence from the product page or have the participant rehearse the presenter's selling points.

Before a street exchange, record who participates, why the interviewer is there, the accepted task, each person's motive and the brand's role. Separate filming assumptions from visible context. A clear later reaction can open the ad; greetings and consent dialogue need not enter the cut. Include preparation, handover and listening time beside the words. If setup exceeds the recipe, revise the premise or flag a longer route. Faster speech and a pasted brand card cannot repair missing context.

Create candidates.json (files reference) with every concept, hook, beat and visual plan.

Step 6. Validate, critique, repair

Run the rule check in strict mode (shape, brand rules, customer words, angle context, references), fix errors and material warnings (the hook and voice rules too, before review), then run the brief-blind opening review and the independent critic from a different model family. Podcast and street dialogue need at least 8/10 on spoken and template_fit in every pass (street also on strategic_fit), with supported claims. Repair at most twice, then report the remaining defect; never relax the rubric or pick the best rejected candidate. Follow validate, critique, repair for the exact checks, critic exits and the optional Jev screen.

Step 7. Deliver the review

Take the strongest eligible concept and validated hook into the recipe's native script shape, with its visual plan. Show up to two viable alternatives in the same review when the angle was open. Do not add a separate pause.

Explain briefly why this promise suits this audience and what the video will show. When research is provisional or essential proof is missing, say so plainly. Keep private quote links and customer data out of remixable review payloads; retain them in research provenance.

Step 8. Learn from decisions and results

Read and update the brand's script history: selected angle id, template and recipe version, hooks, user edits, rejected concepts and reasons. Preserve the user's exact edits. Store standing brand rules through the runtime's normal brand update flow.

Keep user approval, generated quality scores and paid performance separate. A liked script is not a converting ad. When results arrive, attach spend, impressions, audience, placement, objective and window. Learn which angle–format pairs worked without treating every loss as a copy failure.

Quality checks

The selected angle survives, the claim ledger is supported, the script fits the actual recipe, every essential visual can be made, the hook pays off, and the user receives a complete proposal that does not require repairing obvious defects.

Related

• The ad-angle-miner skill — evidence and shared video handoff.
• The watch skill — observe reference footage when its beat map is missing.

Delivery handoff

Include the chosen delivery style, per-beat word counts and speech windows, intentional pauses, recipe conflicts and unverified targets in the existing review. Production uses the same plan and preserves recipe limits. Review finished audio for the complete read, pronunciation, naturalness and sync before declaring delivery fixed.

Carry the brief's pronunciation source, energy progression, emphasized words, pauses and phrases that should sound conversational into the voice-performance task. Word budgets stay provisional until the audio is measured; never use a global speed-up to certify fit.