Analyze video → reusable prompting template

Someone found a video style they love. Your job is to deconstruct it into a reusable prompting template — a formula they can plug any product, person or setting into and get that same style back from seedance-2.0.

The output is not a single prompt. It is a formula file saved into skills/novoads-api/prompting/prompt-library/, built the way seedance-2-ugc.md is built: layers, variables, option banks, rules, and a worked example, so the agent can generate unlimited prompts in that style.

Nothing in steps 1 to 6 touches the API. Frame extraction, transcription and analysis are local work on the user's own file: no calls, no credits. Only step 7, the optional test render, spends anything, and it runs through the two gates in SKILL.md like every other generation.

There is a hosted alternative, and it is not the default. POST /v1/analyses reads an uploaded ad into a structured breakdown (hook boundary, beat timeline, on-screen text, casting, layer zones) in one synchronous call, at a flat fee per analysis that does not move with length. Price it through POST /v1/estimates with {"kind":"analysis"} first. Two reasons it is the fallback and not the front door here: the local path costs nothing and reads the whole runtime, while /analyses defaults to the first 20 seconds (120 is the ceiling, and raising it trades hook precision for coverage). Reach for it when ffmpeg is not installed, when the user wants zone labels rather than prose, or when the frame read has already failed. Offer it with the price attached and let the user choose; never substitute it silently for step 1.

The 15-second constraint is the whole design problem. seedance-2.0 and seedance-2.0-mini take any integer 4 to 15 seconds, and the reference video is usually 30 to 60. Your template has to distil the style into what fits one 15-second clip, and say how to recreate the longer arc across a series of clips.

Dependencies

• ffmpeg / ffprobe — required for frame and audio extraction (step 1). brew install ffmpeg on macOS.
• whisper — optional, for transcription (step 2). pip3 install openai-whisper. If it is unavailable, ask the user to paste the dialogue instead.

Neither is a Novoads dependency. Both run locally on the user's file.

REST key required. A Novoads MCP connector is not a substitute. If NOVOADS_API_KEY is missing or still the placeholder, stop before any generation work and tell the user: "Before continuing, create an API key at https://novoads.ai/dashboard/settings?tab=api and paste it into .env." That holds even when mcp__novoads__* tools are connected and authenticated in the session. Never call mcp__novoads__* tools from this repo's workflows: they are a different surface with different behavior, including the units they quote costs in. Repo installs verify with ./scripts/check-novoads-env.sh; a solo install checks NOVOADS_API_KEY in the environment.

Pack version. Every /v1 response carries X-Novoads-Pack-Version; mention a newer pack at https://github.com/novoads/agent-skills only when that header names a version NEWER than this file's metadata.packVersion — equal or older is nothing to say, and it is never a reason to stop.

Inputs

• Video file (required): path to .mp4, .mov, .webm or similar.
• Style name (optional): what to call the template — car-review, unboxing-hype, skeptic-converted. If they do not give one, name it from what you observe.

Step 1: Extract frames and audio

bash "skills/analyze-video/scripts/extract-frames.sh" \
  "<video_path>" "/tmp/video-analysis" <num_frames>

Frame count by source duration:

Source duration	Frames
Under 10s	8
10–20s	12
20–30s	16
Over 30s	20

Read metadata.txt for duration, resolution and fps.

Step 2: Transcribe the audio

Try, in order:

1. whisper CLI: whisper /tmp/video-analysis/audio.wav --model base --output_format txt --output_dir /tmp/video-analysis
2. Python whisper inline.
3. Neither available → ask the user for the dialogue, or to install whisper.

The transcript is where pacing lives: speech rhythm, filler words, how dialogue interleaves with action. All of that defines the style, and all of it has to survive into the template.

Step 3: Study the reference templates

Read the shipped formulas before you write one. They are the standard your output has to meet:

• seedance-2-ugc.md — 9-layer UGC formula, the richest example
• seedance-2-premium-reveal.md — dark-void reveal, no person
• seedance-2-product-hero.md — elemental product hero, no person
• seedance-2-studio-lookbook.md — studio lookbook with voiceover
• seedance-2-feature-walkthrough.md — feature demo, multi-clip series

And the platform guide, which every one of them defers to:

• seedance-2.md — request fields, the grid, prompt craft, what the estimate flags, the adaptation checklist

Notice what they share:

• They name layers — the structural building blocks of the style.
• Each layer has a pattern: a repeatable sentence shape with {{VARIABLES}}.
• Variables come with option banks, not blanks.
• They state rules that explain why a choice matters.
• They declare a mode — startImageAssetId or referenceAssetIds — at the top.
• They carry a worked example that has been priced live.

Your template has to hit that depth to be usable.

Step 4: Analyze the frames — find what defines this style

Read all the extracted frames. You are not describing one video; you are isolating the transferable pattern.

For every dimension ask: "is this specific to THIS VIDEO — the person, the product, the room — or is it THE STYLE?" Only the style goes into the template. The specifics become variables.

Structure and pacing

• How long is the source? How many distinct beats?
• Which 2–3 beats are essential — the ones without which it stops being this style?
• What is the arc? Hook → demo → proof → verdict, or something else?
• Fast cuts or held shots? How long is each beat?
• Silent beats, or wall-to-wall dialogue?
• The compression question: if you had to carry the whole feel in 3 beats and 2–3 spoken lines, which moments survive?

Camera and framing

• Filming perspective: selfie, propped phone, second operator, screen recording?
• How does framing change between beats — tighter, wider, same angle throughout?
• Is there a signature move that is the style?

Edit style

• Jump cuts, continuous take, time-lapse, split screen?
• Transitions: hard cuts, dissolves, text?
• Recurring motifs: close-up product inserts, reaction face, before/after?

Dialogue and script structure

• Hook format: question, bold claim, mid-action, reaction?
• Scripted, improvised, voiceover, text-on-screen?
• Speech patterns that carry the tone: filler words, sentence length, vocabulary.
• How lines relate to what the hands are doing.

Tone and energy

• 3–4 emotion words for the vibe.
• Energy arc: builds, flat, peaks then drops?
• Relationship to the viewer: friend, expert, skeptic, fan?

Lighting and technical quality

• Light source and direction: natural, ring light, moody, blown out?
• Phone or polished? Which technical "flaws" are load-bearing?
• Audio character: phone mic, lapel, voiceover, room tone?

What makes this style DIFFERENT

The most important pass. After cataloguing everything above, name the 2–3 things that separate this from a generic UGC clip or a generic product review — the pacing, the hook format, the way the product enters frame, the edit rhythm. Those become the core of the template; everything else is scaffolding.

The 15-second plan

Before building anything, map the style onto one clip:

• What is the minimum viable version? Which beats are essential, which are nice to have. Fifteen seconds has to carry the thing that makes someone say "oh, that kind of video".
• Does it need a series? If the power is in a narrative arc or a feature rundown, it needs 2–3 clips. If it is a vibe or a single moment, one clip is enough.
• How many spoken lines fit? Delivery measures 2.0 words per second (~13 characters a second, spaces included) after ~0.5s of leading silence. Fifteen seconds is 2–3 short sentences, and the slack is what leaves room for a silent beat. Count the source's lines and keep the ones carrying the voice.
• What is the beat skeleton? 15 seconds is 2–3 beats: hook → core moment → kicker.

Which mode the style wants

Every formula in this library declares one, because they are mutually exclusive on the API and the choice follows from the style:

If the style…	Mode	Why
opens on the product itself, held up or sitting on a surface	startImageAssetId	the product photo is literally the first frame
builds a scene the product was never photographed in — a void, a splash, a studio set	referenceAssetIds	the model has to composite it, not animate a flat photo
holds one person across a series of clips	referenceAssetIds	@Image1 the product, @Image2 the person — Seedance re-casts on every cut, and a repeated description does not hold a face

Write the answer into the template. A formula that leaves the mode open produces prompts whose @Image1 tokens point at nothing.

Step 5: Build the template

Create a self-contained markdown file. Someone should be able to read it and write prompts in this style without ever seeing the source video.

Template structure

# [Style name] — Seedance 2.0

**Use when:** [the kind of video this produces]

**Model guide:** read [seedance-2.md](../novoads-api/prompting/prompt-library/seedance-2.md) first for the request fields, the grid,
and the platform rules.

**Mode:** `startImageAssetId` or `referenceAssetIds` — say which, and what each `@ImageN`
slot holds. The two are separate modes and a body carrying both is a `400`.

## What defines this style

[2–3 paragraphs of theory. This is what lets a prompt writer make good variable choices
instead of filling blanks.]

## Things to know before you write a word

[The route-specific traps. If the style speaks, say that the line is rendered and
lip-synced in this same call and that gate 1 applies. If it is silent, say that silence has
to be declared in the prose.]
mandatory.]

## The structure

[The layers of THIS style. Do not force-fit the 9-layer UGC model. Five layers, twelve
layers — let the video decide.]

## Layer-by-layer formula

### Layer N: [Name]

[What this layer does and why it matters here.]

**Pattern:**
\```
[The repeatable sentence shape with {{VARIABLES}}]
\```

| Variable | Options | Notes |
|---|---|---|
| `VARIABLE_NAME` | option 1, option 2, option 3 | [guidance] |

### [... more layers ...]

## Beat structure (one 15-second clip)

[The 3-beat framework: hook, core, kicker. Which beats speak and which are silent. Two to
three spoken lines total.]

## Multi-clip strategy (if applicable)

[How to split across 2–3 clips, what each one covers, and what holds identity across them:
the same `referenceAssetIds` in every call plus the actor tag repeated verbatim.]

## Tone and pacing guide

[Energy, speech patterns, rhythm, with a pacing-cue bank specific to this style.]

## Technical specs

[Lighting, camera quality, audio character — including the flaws that make it authentic.]

## Complete template

[One copy-paste block for ONE 15-second clip, every variable marked {{PLACEHOLDER}}.
Max 3 beats, max 2–3 spoken lines. This is the unit.]

## Example prompt

[The template filled in for a DIFFERENT product, person and setting than the source video.
Price it at POST /v1/estimates before shipping it.]

## Adaptation checklist

[The style-specific checks, then the standard ones from seedance-2.md.]

## Generating from this template

[The call sequence — see step 7 of analyze-video for the block to paste here.]

Rules the template has to teach

These are not style preferences. Each one is a defect found in a shipped formula, and a template that omits them manufactures the same defect in every prompt written from it.

• Every prompt is one clip of 4 to 15 seconds. If the style needs more, the template ships a multi-clip strategy — never a longer prompt.

• Keep prompts between 100 and 260 words. Shorter prompts produce vague results; longer ones overwhelm the model and cause it to lose focus on key details. Do not pad to hit the floor and do not cut a beat's framing to stay under the ceiling.

• Variables are curated choices, not blanks. "Any lighting" is useless. "Natural window light, overhead kitchen light, golden-hour balcony light" is a decision the writer can actually make.

• Declare the mode and address references as @Image1, @Image2, … The tokens resolve positionally against referenceAssetIds in the order the array is sent, and a token pointing past the end of the array is refused before the charge. If the style uses a start frame instead, say so and use no tokens at all.

• If the style is silent, say so in prose — a silent product film with no spoken dialogue, silent b-roll. Seedance renders audio from the prompt, so a film that never declares silence can come back with an invented voice on it. Do not put on-screen text in double quotes and call it done — the model reads a quoted string as a line to speak.

• No forbidden words: cinematic, professional, stunning, 8k, studio, perfect. Nothing on the API rejects or reports them — this is craft advice, and the reason to drop them is the render. Replace one with the real thing: the light source, the surface, the flaw.

• Prose, never a bulleted prompt. A run of Label: value pairs or - lines comes back rendered as literal text on screen. Use timestamps — [00:00], [00:05] — for multi-beat pacing instead.

• One primary action per shot, with two or three comma-joined cues on it. A second action chained with then / and then / followed by renders as a smear. Split it into two shots.

• Repeat the actor tag verbatim. the same woman resolves to nobody; identity does not carry across a cut.

• The example must use different content than the source. If the source was a woman reviewing a serum in her bedroom, the example is a guy reviewing a protein bar in his kitchen. That is what proves the template generalises.

• Dialogue has to fit. Count the words: 2.0 per second measured, so a D-second clip holds about 2.0 × (D − 0.5) once the leading silence is paid for. If the source talks fast, say so in the template and keep each line punchy.

• No credit numbers anywhere in the template. Prices come from a live POST /v1/estimates in the session that spends them. A template that quotes a number teaches the next agent to skip the call.

Step 6: Save, register and present

1. Save to skills/novoads-api/prompting/prompt-library/seedance-2-<style-name>.md.
2. Add a row to the style directory in seedance-2.md so the new formula is reachable — user goal, file link, key trait, matching the rows already there. A formula nothing points at is a formula nobody reads.
3. Summarise in chat: the style you identified, its layers, what makes it distinct, and the path you saved it to.
4. Ask: "Want me to price a test prompt from this template and render it, to prove it works?"

If they say yes, write the prompt for a different product, person and setting than the source — that is the actual test — and run step 7.

Step 7: Render a test clip (optional, and it costs credits)

The template is the deliverable; this step only proves it. It runs the full sequence from SKILL.md, and both gates apply.

1. Upload the product photo if the prompt references one:

   curl -sS -X POST https://api.novoads.ai/v1/uploads \
     -H "Authorization: Bearer $NOVOADS_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{"contentType":"image/jpeg","sizeBytes":248193}'
   

   then PUT the raw bytes to the uploadUrl it returns, sending back exactly the headers it returned. The assetId is durable and reusable across calls, models and sessions — upload once and keep the id.

2. Price it — gate 2. POST /v1/estimates with kind: "video", the prompt, the model and durationSeconds. It charges nothing and it is the only source of a price. It also returns an advisory warnings array of craft notes (verified live 2026-08-04) — read them, but they are substring matches that false-positive, and the checklist above is still the real quality gate. Show the number and get a yes before spending.

3. Confirm the spoken line — gate 1, if the clip speaks. Numbered beats, word count against the duration, the language you are sending, an explicit yes. Approving the template is not approving the sentence.

4. Generate:

   curl -sS -X POST https://api.novoads.ai/v1/videos \
     -H "Authorization: Bearer $NOVOADS_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "seedance-2.0",
       "prompt": "<the filled-in prompt>",
       "durationSeconds": 15,
       "aspectRatio": "9:16",
       "language": "en",
       "referenceAssetIds": ["<assetId>"],
       "productId": "<uuid>"
     }'
   

   Swap referenceAssetIds for startImageAssetId if that is the template's mode. Never both. Returns 202 with jobId and creditsCharged.

5. Poll GET /v1/generations/{jobId} every 15 seconds until a terminal status — succeeded, failed, blocked or canceled. Not until succeeded: a loop waiting for success never returns on a job that died. seedance-2.0 usually takes 3 to 8 minutes, most often around 5; seedance-2.0-mini 2 to 3.

6. Download GET /v1/generations/{jobId}/watch, save under outputs/<descriptive-subfolder>/, and open the folder so the user can watch it.

Draft on seedance-2.0-mini first. It is the same grid, the same fields and the same prompt at half the price, back in 2 to 3 minutes — and since the assetId is durable, the final render on seedance-2.0 reuses the same upload. A template validated on mini is a template validated.

Paste this block into every template you generate, so the formula carries its own call sequence:

## Generating from this template

1. Upload the product photo: `POST /v1/uploads` → `assetId` (durable, reuse it).
2. Price it: `POST /v1/estimates` with `kind: "video"`, `model`, `durationSeconds`,
   `prompt`. It charges nothing, it is mandatory, and it is the only source of a price — show the number out loud
   before spending.
3. Confirm the spoken line with the user if the clip speaks (`SKILL.md` gate 1).
4. Generate: `POST /v1/videos`

\```json
{
  "model": "seedance-2.0",
  "prompt": "<your filled-in prompt>",
  "durationSeconds": 15,
  "aspectRatio": "9:16",
  "language": "en",
  "referenceAssetIds": ["<assetId>"]
}
\```

   `startImageAssetId` instead if that is this formula's mode — never both, that is a `400`.

5. Poll `GET /v1/generations/{jobId}` every 15s until a terminal status, then download from
   `GET /v1/generations/{jobId}/watch`.

Every clip in a series is its own call and its own charge. Five generations per
organization may be in flight at once.

Related files

• clone-video-ad/SKILL.md — the sibling: same analysis, but the output is a generated video for the user's product instead of a template file.
• scripts/extract-frames.sh — frame and audio extraction, shared by both skills.
• seedance-2.md — the platform guide every generated template defers to. The other two live video models carry their own grids, and you need one the moment a template targets them rather than Seedance: sora-2.md when the first spoken word has to land immediately (it measured no leading silence, where Seedance front-loads 3–5s), veo-3-1.md when the user names Veo or the shot has to evolve over its runtime. Neither takes referenceAssetIds.
• ugc-selfie-style.md — the cross-model UGC guide. Its Core principles transfer to Seedance; its per-model formulas do not — never port one across by find-and-replace.
• ../../SKILL.md — the call sequence, the two gates, polling, download.
• ../../reference.md — every endpoint, field, limit and error code.

File map

skills/novoads-api/
├── SKILL.md                                  ← router: decision tree, gates, full sequence
├── reference.md                              ← endpoints, fields, limits, errors
└── prompting/
    ├── guide.md                              ← marketing brief → API
    ├── brand-voice-starter.md                ← template to copy into MASTER_CONTEXT.md
    ├── analyze-video/
    │   ├── SKILL.md                          ← THIS FILE — video → reusable template
    │   └── scripts/extract-frames.sh         ← ffmpeg frame + audio extraction
    ├── clone-video-ad/
    │   └── SKILL.md                          ← video → adapted video for the user's product
    └── prompt-library/
        ├── seedance-2.md                     ← Seedance 2.0 platform guide (read first)
        ├── seedance-2-ugc.md                 ← 9-layer UGC formula
        ├── seedance-2-premium-reveal.md      ← dark-void premium reveal
        ├── seedance-2-product-hero.md        ← elemental product hero
        ├── seedance-2-studio-lookbook.md     ← studio lookbook with voiceover
        ├── seedance-2-feature-walkthrough.md ← feature walkthrough demo
        ├── ugc-product-selfie.md             ← image formulas (product selfie)
        ├── product-showcase.md               ← image formulas (product showcase)
        ├── influencer-recreation.md          ← image formulas (likeness)
        ├── character-sheet.md                ← image formulas (character consistency)
        ├── character-sheet-gpt-image-2.md    ← the same on gpt-image-2
        ├── nano-banana.md                    ← image formulas (Nano Banana Pro)
        ├── sora-2.md · veo-3-1.md            ← the other two LIVE video models' grids
        └── kling-3.md · ugc-selfie-style.md  ← Kling 3.0 is on the API as `kling-v3-pro` (2.38.0); the selfie guide is cross-model

New formulas you write land in prompt-library/ beside the Seedance files, and get a row in seedance-2.md's style directory.