Novoads REST API

Spend your effort on the prompt and on confirming the spoken line. The HTTP is mechanical and is written out below; the prompt is where the render is won or lost.

Configuration

• Base URL: https://api.novoads.ai/v1 (host overridable with NOVOADS_BASE_URL — host only, you append /v1).
• Auth: Authorization: Bearer $NOVOADS_API_KEY, read from .env at the repo root. The key is novo_ plus 64 hex.
• Check: ./scripts/check-novoads-env.sh
• Never print API keys, commit .env, or paste a key into MASTER_CONTEXT.md.

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.

If the key is missing or the API returns 401 or 403

Run ./scripts/setup.sh, or tell the user to create a key at https://novoads.ai/dashboard/settings?tab=api. Prefer the editor: ask them to paste the key inside .env rather than into chat. If they do paste it in chat, write .env for them, confirm "saved to .env" without repeating the key, and tell them chat history may retain secrets, so the key is worth rotating if that transcript could be shared.

A 401 means the key is wrong, revoked, or from another account. A 403 with error.details.reason of plan_required or subscription_inactive means the key is fine but the account has no live subscription. Those are different problems and you should say which one it is.

Signup link

If the user does not have a Novoads account, give them this link at any point in a session, in any context where it is useful:

https://novoads.ai/?utm_source=claude-code&utm_medium=github&utm_campaign=skill-pack

There is no free tier and no trial: every plan is a paid monthly subscription, published at novoads.ai/pricing. Never call it free.

After the connectivity check passes

When ./scripts/check-novoads-env.sh returns 200 and the user has not yet said what they want, offer starting points in chat and stop. Nothing is printed by a script for you: a banner cannot see their product, their references/ folder, or their language, so it suggests nothing.

1. A UGC video built from their own product photo. Ask them to drop it into references/. Route: seedance-2.0 plus startImageAssetId.
2. The same idea on seedance-2.0-mini first, to get the prompt right before the final render. Price both at POST /v1/estimates and show the difference rather than quoting one from memory.
3. A static image ad, when what they need is a still and not a clip. Route: gpt-image-2 for heavy text or a mimicked UI, nano-banana-pro for a photoreal scene.

Offer, do not choose. A first render fired on a guess is a charge the user did not ask for.

Read order

1. Repo root MASTER_CONTEXT.md when present: brand voice, default product, accumulated decisions. It carries no prices — that is deliberate, see gate 2.
2. This file. It is the router and it covers the full call sequence.
3. List references/ at the repo root before you ask the user for a photo. It is where they keep product shots, actor stills, and style boards, and it is gitignored, so the files are theirs and are not in this skill's folder. A product photo found there becomes startImageAssetId (video) or an entry in referenceAssetIds (images). An empty references/ is the normal state of a fresh clone, not a blocker — see below before you ask for anything.
4. MANDATORY before composing any prompt: the prompting/prompt-library/ file for the route you picked in the decision tree. The libraries carry the craft; every HTTP detail comes from this file and reference.md, which win whenever the two disagree.
5. reference.md when you need a field you do not see here, or when you hit a status code you want to branch on.

Decision tree

The user wants	Route
A UGC video: a person talking to camera about a product	seedance-2.0. Read seedance-2-ugc-v2.md for the shape — how many beats, how many API calls, which reference mode — then seedance-2-ugc.md for the craft of writing each layer. v2 is one render with four talking beats in one location; it is the default and costs about half what stitching costs
The user says "a start frame for each scene", or asks for several scenes	That phrasing commits you to one call per scene (startImageAssetId), which forecloses multi-beat cuts inside a single render — the two modes are mutually exclusive. Say so, and offer the one-shot alternative from seedance-2-ugc-v2.md before generating anything
B-roll cutaways, burned captions, background music, or variations of a finished ad	Not part of making the ad. Deliver the base video first, then offer these as a separate pass, each owned by its own skill: broll-overlay (needs the base and its transcript from POST /v1/transcripts — no local install; it overlays — the base audio keeps running and the final duration is unchanged — it does not extend), POST /v1/captions, then music-mix last
The same thing, but cheap, to test a prompt before committing	seedance-2.0-mini, same grid, same formulas, half the price and back in 2–3 minutes. The draft-then-finalize loop is Mini-draft tier in seedance-2.md
A clip longer than 15 seconds, or seedance-2.5 by name	seedance-2.5 — the only model on this API that renders past 15s (any integer 4 to 30). Same six aspect ratios, same 9 referenceAssetIds, same audioEnabled, same 4,000-character ceiling, so the Seedance formulas apply unchanged; write for the longer runtime rather than padding a 15s script. Two things differ: it renders 480p and 720p only — no 1080p, no 4k, on either provider — and nobody has timed one, so quote its wait as unknown. Price it explicitly: at the same length it is dearer than seedance-2.0, and thirty seconds of it is the most expensive single call on this API
A premium product reveal: dark void, no person, text narrative	seedance-2.0 + seedance-2-premium-reveal.md
A product hero: elemental effects, splash or mist, no person	seedance-2.0 + seedance-2-product-hero.md
A studio lookbook: polished, voiceover, multi-look	seedance-2.0 + seedance-2-studio-lookbook.md
A fast-paced feature walkthrough	seedance-2.0 + seedance-2-feature-walkthrough.md
A fast vertical clip with no dialogue requirement	omni-flash, and read shared/skills/gemini-omni-flash/prompting/guide.md first — its 20,000-character prompt ceiling is the roomiest of any video model here, where the rest sit at 4,000, and its grids are narrower than Seedance's (durationSeconds 4/6/8/10 only, aspectRatio 9:16 or 16:9 only, resolution read from GET /v1/models). Since API 2.30.0 it also takes reference images, one reference video and, where the server publishes them, a first frame, a last frame and a seed: see Start frame or references below
A talking clip whose first word must land immediately	sora-2, and read sora-2.md first. Measured here at no leading silence at all, where Seedance front-loads 3–5s of it. Grid is durationSeconds 4/8/12 only, aspectRatio 9:16 (default) or 16:9, startImageAssetId yes, referenceAssetIds no
Veo 3.1 by name, or a shot that has to evolve over its own runtime	veo-3.1, and read veo-3-1.md first. Grid is durationSeconds 4/6/8 only, aspectRatio 9:16 (default) or 16:9, startImageAssetId yes, referenceAssetIds no. Nothing in this repo has measured a Veo render — quote its wait as unknown rather than borrowing Seedance's
A video that starts from a specific photo	any video model but omni-flash plus startImageAssetId, which animates that image as the first frame. On omni-flash that field is a reference image the model may place anywhere in the clip; for a true first frame use firstFrameAssetId where the server publishes it
The picture the user handed over is itself the shot: a UI screenshot, a hero layout, a flat-lay, key art, and they want it to MOVE ("animate this", "make the cards pop in")	image-to-motion, which is seedance-2.5 plus startImageAssetId and carries the motion vocabulary, the beat template and the text-fidelity clauses. The split is whether the image is preserved and set in motion (there) or seeds a scene that did not exist (here)
A video built from several photos: the actor and the product, a wardrobe, a setting	seedance-2.0 or mini plus referenceAssetIds — up to 9 images, composited rather than animated, addressed in the prompt text as @Image1…@ImageN in the order you send them. Seedance, or omni-flash since API 2.30.0 (its limit is the field's maxItems in the OpenAPI document; sora-2 and veo-3.1 have no such field), and never alongside startImageAssetId: they are separate modes and a body carrying both is a 400
The same person to hold across several clips of a series	pass that person's photo in every clip's referenceAssetIds and repeat the actor tag verbatim. Seedance re-casts on every cut, so a repeated description alone does not hold a face; see seedance-2-feature-walkthrough.md
A reference video turned into a reusable template: "make videos like this", "deconstruct this"	read prompting/analyze-video/SKILL.md. Frames and transcript are extracted locally with ffmpeg and Whisper, and the output is a new formula file in prompting/prompt-library/. Nothing is charged until the optional test render at the end
One specific ad cloned for their own product: "make this ad but for my product"	read prompting/clone-video-ad/SKILL.md. The same local analysis, but the output is a rendered clip and both gates apply. A source longer than the chosen model's ceiling becomes a series, held together by passing the same referenceAssetIds to every clip — there is no in-place restyle of footage on this API (a reference video only guides a new render). seedance-2.5 reaches 30s in one call, so a source that used to need two clips may now need one; price both shapes before you pick
A static ad with heavy text or a mimicked UI	gpt-image-2
A photoreal still: a person, a product in a scene	nano-banana-pro
A different look on a still, or a second opinion on one	reve-2.1
The dashboard's Image to Ad, Product to Ad or Upscale tool by name, or a still upscaled to 4K	POST /v1/images with preset (image-to-ad, product-to-ad or upscale, API 2.40.0) on one photo in imageAssetId, and no model or prompt: the server writes the prompt. Quote it with the same preset on POST /v1/estimates; it can answer running like any image. Fields: reference.md
A Pixar-style 3D animated ad, at ANY length	read skills/pixar-ad/SKILL.md: storyboard on gpt-image-2, animate each beat on seedance-2.0 + startImageAssetId, narration into the gaps from POST /v1/voiceovers, stitch with ffmpeg. Every animated ask routes there, including "a quick 15 second one" — there is no single-call tier. Nothing on the API rejects, checks or comments on a stylized prompt, so that skill is the only thing that will tell you whether the beat works
A claymation / stop-motion clay ad	read skills/claymation-ad/SKILL.md, same shape over 8 beats
Captions burned onto a finished MP4	Two real paths — offer both. POST /v1/captions: one call, 30 preset styles, no local dependencies, costs credits and returns only a new MP4 (never an SRT or the caption text). Or the caption-video skill (shared/skills/caption-video/SKILL.md): free, any style you can write, and it gives you a Whisper transcript you can hand-correct — but needs Whisper, HyperFrames and an ffmpeg alpha composite locally. See Burned-in captions below. A clip rendered with audioEnabled: false can only go the local route
Meta image-ad creatives from a brief or a template	read shared/skills/image-ad-prompting/OVERVIEW.md first, then chatgpt-image-ad or nano-banana-image-ad
To reverse-engineer an existing image ad into a reusable template	the clone-image-ad skill
A YouTube thumbnail	the generate-youtube-thumbnail skill
B-roll, an ambient product clip, a scene	There is no b-roll endpoint. Generate a silent clip: omni-flash, or seedance-2.0 with the word silent or b-roll in the prompt
Kling	kling-v3-pro where GET /v1/models lists it (API 2.38.0): any integer 3 to 15 seconds, 16:9, 9:16 or 1:1, one resolution, audioEnabled on or off, a text prompt or startImageAssetId as the first frame, and since API 2.44.0 lastFrameAssetId as the last frame beside it, no referenceAssetIds. Read the grid from GET /v1/models. kling-3.md is vendor prompt craft whose field names belong to another API, so take the body from reference.md. Nothing in this repo has measured a Kling render: quote its wait as unknown
A library actor holding their product, or moving like a clip they have (a dance, a gesture)	Both cast an actor from GET /v1/actors, never an uploaded face. Holding the product: productSwap: { actorId, productAssetId } on a Seedance-family POST /v1/videos (API 2.42.0), the product photo from POST /v1/uploads, not combinable with startImageAssetId, priced with productSwap: true on the estimate (reference.md). Moving like a clip: POST /v1/videos/animate-actor (API 2.43.0) with the driving clip from POST /v1/uploads, whose measured length is the output's length and the price (reference.md)
A video's background removed, one rendered here or their own	POST /v1/background-removals (API 2.41.0) with a jobId or an uploaded assetId, up to 30 seconds, a longer source refused and charged nothing (reference.md)
To edit an existing MP4 they already have	Not this skill, except captions (row above) — and note that POST /v1/captions accepts an uploaded assetId, so burning subtitles into their own file is a supported first-party call, not just a local one. Everything else (trims, cuts, overlays, music) is out of scope. Say so
To publish the result as an ad on Meta or TikTok	Not this skill. The output is a file. The meta-ad-builder skill takes it from there

Prefer the shortest path. If one model answers the request, do not build a pipeline around it.

Step 0: classify before you call anything

Refusing is a successful result. If the request is to edit a file the user already has, or to publish to an ad platform, say the skill does not do that and stop. Do not improvise a pipeline out of the generation endpoints.

The two gates

Two separate approvals stand between a request and a charge, and neither implies the other. Approving a concept is not approving a sentence, and approving a sentence is not approving a spend.

Gate 1 — the spoken line (MANDATORY for any video with dialogue)

Seedance renders the audio and the lip-sync in the same call, so the line inside the double quotes is what the actor says, out loud, in the finished video. It cannot be changed afterward without paying for the render again.

1. Extract the dialogue from the drafted prompt and show it on its own, separate from the visual description.
2. Spell any invented brand name phonetically inside the quoted line — see the rule below. Do this before you present the line, and show the phonetic form in the block, because it is what the model will be sent.
3. Present it as a numbered list with beat labels (hook / show / demo / verdict, or similar). Mark silent beats (silent beat — no dialogue).
4. Count the spoken words, state the target duration, and say whether it fits at a natural pace.
5. State the language you are going to send, because that is the language the ad is rendered in.
6. Ask for approval explicitly. Never infer it from an earlier yes about tone, template, or cost.

Use this structure:

📝 Dialogue script (please confirm before I generate)

  1. [HOOK]    "Bro. BRO. Look what just showed up."
  2. [SHOW]    "That colorway? Insane. Like, who greenlit this?"
  3. [DEMO]    (silent beat — thumb brushing the suede, small nod)
  4. [VERDICT] "I'm wearing these to the gym tomorrow. You have to see them in person."

Total spoken words: ~28  |  Target duration: 15s  |  language: en  |  Fits at natural pace: ✅

Approve this dialogue? (yes / edit / rewrite)

If they say edit, revise and re-present the block until they approve. The gate applies to every flow where the model speaks. Skip it for silent flows: product hero, premium reveal without voiceover, and images.

Re-run the gate whenever the words change, including when you change them yourself for the rule below. An approval covers the sentence that was approved, not its successor.

Invented brand names get a phonetic spelling in the quoted line

The gate approves text; the model renders speech. A brand name that is a real word survives the trip. A coined or portmanteau name may not: on 2026-08-02 seedance-2.0 rendered Novoads as "Nuvenov's" — unrecognisable as the brand — while speaking the other 16 words of the line verbatim. Approving the sentence did not approve how it would be said.

The rule: write the name as hyphenated syllables inside the double quotes.

… and says: "I kept saying AI ads look fake until I made one on NO-vo-ads and nobody could tell."

Use it for names that are invented, run two words together, or that a reader would have to guess at. Leave ordinary words alone — Nike and CeraVe do not need it.

This form is validated in en only. See the es limit below before you reach for it in another language — the same brand fails there for a different phonetic reason, and the English spelling does not fix it.

What the A/B actually showed (job 6329f29a… vs ff69d118…, single variable, every other byte identical, pass criterion fixed in writing before the render): NO-vo-ads came back transcribed as "Novo ads" — recognisable — and the feared failure did not happen: the model did not read the hyphens aloud or spell the name out.

Its limits, which are as much a part of the rule as the rule:

• n = 1. One brand, one model, one language. The fix is validated on seedance-2.0 in en only — untested on sora-2, veo-3.1 and omni-flash, and untested in es/pt, where hyphenated English syllables may read very differently. Do not present it as a general fix.
• es needs a DIFFERENT spelling, not the English one. A Spanish render shipping the brand plain came back as "NoBots" — unrecognisable. Spanish has no /v/–/b/ contrast — ⟨v⟩ and ⟨b⟩ are one phoneme — so no spelling buys back the /v/, and what actually broke was the tail: the o+a hiatus collapsed and a syllable vanished, vo-ads → bots. Hyphenating syllable boundaries is all NO-vo-ads does, and it addresses none of that.
  • The es form is Novo Ads — two words, with a space. Validated by A/B (6ee797e0 vs d8aaee21, single token changed, pass criterion fixed in writing beforehand): transcribed "Novo Ads", recognisable. The orthographic word break restores the syllable the hiatus ate; it does not try to restore the /v/, because nothing can.
  • Provisional, and the failure mode is nasty. The identical payload was 1-for-2. On the bad take Seedance stuttered — one clause spoken twice, verified as real speech and not a decoder artifact — and Novo collapsed into the preposition, leaving the ordinary Spanish phrase "no ads". That is worse than "NoBots": a listener hears no brand at all rather than a mangled one. Transcribe every es render. A clean take is not evidence the next one is clean.
  • pt is untested in both directions.
• It lands as two words — "Novo ads", not "Novoads". That matches what sora-2 produced unaided and a listener will recognise it, but it is not a perfect single-token rendering.
• It fixes pronunciation and nothing else. Leading silence went 3.71s → 3.16s, i.e. unchanged. The dead-air problem is a duration problem; see Script length → duration.
• Seedance re-cast the actor between the two renders, so the A/B held the prompt constant, not the performer. Some part of the delta could be voice-casting luck. Enough to act on, not enough to call proven.

Verify it in the render. The rule is a prompt fix with a measured result, not a guarantee — the video QA step in §7 is what tells you the name actually came out right this time.

Gate 2 — the cost estimate (MANDATORY)

Never state a credit cost from memory, and never generate before showing the user a number that came from a live call in this session. There are no rate tables in this repo, in MASTER_CONTEXT.md, or in the logs. Prices come from POST /v1/estimates at call time, and that is the whole policy.

POST /v1/estimates spends nothing and returns:

{ "credits": 3.2, "balance": 100, "sufficient": true }

When it is short it also returns shortBy and topUpUrl.

/estimates also returns warnings, and they are advice, not verdicts

POST /v1/estimates runs craft rules against your prompt and returns them in a warnings array (verified live 2026-08-04 against spec 2.6.0). An earlier version of this file said the field did not exist and that nothing on the API reads a prompt for quality. Both were wrong. Each entry is { "rule": "...", "message": "..." }, and the message usually quotes the exact substring that tripped it:

{ "credits": 7, "balance": 860.1, "sufficient": true,
  "warnings": [
    { "rule": "label_without_hold",
      "message": "Mentions a label, package or screen with no labelHold clause. … (found in your prompt: \"bottle\")" }
  ] }

Rules seen live: no_spoken_line, missing_actor_descriptor, label_without_hold, chained_motion, no_aspect_ratio. Those are video craft rules. Image prompts are read too, by rules of their own — banned_polish and blank_label observed on a kind: "image" estimate against deployed spec 2.19.0 (verified live 2026-08-12); an earlier probe on 2026-08-04 returned no warnings key for images, which described that deployment, and the advice here to treat image prompts as unlinted is retired. A kind: "caption" estimate has no prompt to read and returns none either.

They are purely advisory. None of them refuses a generation, none changes the price, and a prompt that trips every one of them renders exactly like one that trips none. /estimates is the only endpoint that runs them — POST /v1/videos and POST /v1/images do not, and their responses carry no warnings field.

They produce false positives, and you are the one who has to catch them. Both of these were reproduced live on 2026-08-04:

Prompt	Rule that fired	Why it was wrong
"She turns her laptop screen toward the camera to show the dashboard…"	label_without_hold — (found in your prompt: "screen")	The rule protects printed text on physical packaging. A software product has no label to preserve, and pasting in the labelHold clause would tell the model to hold a package that is not in the shot.
"He says: 'I tried everything for the rust. Then a friend told me about this.'"	chained_motion — (found in your prompt: "Then")	The match is inside a quoted spoken line. "Then" is dialogue, not a second motion instruction — there is exactly one action in the shot. Splitting it would split the sentence the actor says.

Both rules are substring matches. They cannot tell a physical package from a UI, or narration from stage direction.

So handle them like this:

1. Read every warning. They catch real mistakes — a missing spoken line on a Seedance render is money thrown away, and label_without_hold on an actual product package is the single most expensive prompt error in this repo.
2. Judge each one against what your prompt actually says, including where the matched substring sits. Quote the rule and your reasoning when you decide.
3. Never apply a suggested fix blindly. The fix text is a generic clause; pasting it into a prompt it does not fit makes the render worse, not better.
4. Never silently drop one either. If you are overriding a warning, say so to the user in one line — "the label_without_hold warning matched the word 'screen', but this is a SaaS dashboard with no printed label, so I am not adding the clause" — so the call is visible and reversible.

The prompt libraries under prompting/ are still the real quality gate. These warnings are a cheap second opinion collected on a call you were making anyway; they are not a substitute for composing against the formula file, and they say nothing at all about whether the idea works.

The estimate body is not the generate body. It takes only the fields that move the price, plus a kind discriminator, and it is strict: any extra key is a 400.

Estimate accepts	Video	Image
required	kind: "video", prompt	kind: "image", prompt
optional	model, durationSeconds, language, resolution	model, numImages, language, and from API 2.36.0 quality, outputFormat, resolution (as on POST /v1/images)

resolution belongs here because it moves the price — on seedance-2.0 it is the difference between the base and roughly five times it. Send the resolution you actually intend to render, or the quote prices a cheaper video than the one you generate. See the resolution section below for the ladder and for the seedance-2.0-mini trap. (Verified live 2026-08-04.)

aspectRatio, startImageAssetId, referenceAssetIds, and productId do not belong here. They do not change what you pay, and sending one is a rejected request. There is no styleFamily field — not here and not on a generation call; it was removed from the API and sending it is a 400 Unrecognized key.

curl -sS -X POST https://api.novoads.ai/v1/estimates \
  -H "Authorization: Bearer $NOVOADS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"kind":"video","model":"seedance-2.0","durationSeconds":12,"language":"en","prompt":"..."}'

Pass model explicitly. It defaults to seedance-2.0. At the same length the video schedules span more than 10x across the set, and more than 28x across every cell the API publishes once length and resolution are counted too — both figures are derived by the API and printed in the model field's own description, and both moved when seedance-2.5's thirty-second grid landed. Image schedules span more than 3x. The per-model prompt ceiling is enforced against whichever model you name. Pricing the wrong model is a quote that disagrees with the invoice.

It runs the same access checks and the same structural validation the paid call runs, which is why it is worth calling every time and not only when you are unsure:

• It is the one endpoint with an opinion about the prompt — the warnings array above. POST /v1/videos and POST /v1/images have none and return no such field. Compose against the formula file before you price: the warnings are a substring-matching second opinion, not a review, and a prompt that trips nothing can still be a bad prompt.
• What it does still refuse, at no charge: a malformed body — it is strict, and any key that does not move the price is a 400 — and a prompt longer than the named model's character ceiling (4,000 for all three Seedance variants, 20,000 for omni-flash; name no model and it is judged as seedance-2.0).
• On the image arm its length check is looser than the generation's. The estimate's request schema caps prompt at 50,000 characters whatever image model you name, while POST /v1/images caps gpt-image-2 at 32,000 and reve-2.1 at 4,000 (deployed spec 2.16.0, verified 2026-08-08). So an image prompt between the two prices cleanly and is then refused at generation. Only nano-banana-pro, at 50,000, has the two numbers agree.
• A quote it returns cannot disagree with the invoice, with two exceptions worth knowing: it never sees aspectRatio, the asset fields, or productId, and it skips moderation, which the paid call runs. A clean estimate can still come back 422 at generation — moderation is the only thing left that refuses a prompt for what it says.

Multiply before you show. N variations is N charges. Show the per-call number, the count, and the total.

Warn when the total exceeds the balance. balance comes back on the same response: if the batch total is larger, say so before asking for a yes, and quote shortBy and topUpUrl when the estimate provides them. sufficient is a snapshot, not a reservation — another session or a renewal can move the balance between the quote and the call.

Show credits, the count, the total, and balance. Get a yes. Then generate.

Choices you make out loud, and choices you infer

Ask, every time:

• Seedance tier, once per workflow. Before the first Seedance video call: "Use default seedance-2.0, seedance-2.0-mini (half price), or seedance-2.5 (the only one that goes past 15s, and the dearest)?" No preference means seedance-2.0. Whichever they pick goes into the estimate, so the quoted number is the one they pay. Ask before you assume the script needs 2.5 — a 30-second ad is a format decision, not a longer version of a 15-second one.
• How many variations, for every prompt. Default 1. N variations means N identical calls — there is no batch parameter — and the results come back as a numbered list so they can compare and pick.

Infer, and state what you inferred rather than asking:

• aspectRatio: default 9:16 for anything headed to Reels, TikTok, Stories, or a vertical feed. Seedance defaults to 16:9, and a landscape ad is a wasted render; omni-flash, sora-2 and veo-3.1 already default to 9:16, and images default to 1:1. Go landscape only when the user asks. Seedance also accepts 1:1, 4:3, 3:4, and 21:9; omni-flash, sora-2 and veo-3.1 accept only the two.
• resolution (only on a model whose resolutions in GET /v1/models lists more than one value: seedance-2.0, seedance-2.5, from API 2.34.0 seedance-2.0-mini and veo-3.1, and omni-flash where the server lists more than 720p): leave it at the model's defaultResolution, and never raise it silently. On the Seedance models it is the one output-shape field that multiplies the bill (not on veo-3.1, one flat price at every tier) — on seedance-2.0, 1080p is ≈2.5x the base and 4k ≈5x — so going above 720p is a spend decision, not a quality preference, and it belongs in front of the user with a fresh estimate attached. 480p is now ≈half the base rather than the same price (family reprice, 2026-08-07), so it is a real draft tier and worth offering when a render is a rehearsal. seedance-2.5 takes 480p and 720p and nothing above. omni-flash's set depends on the server (it may list 360p, 720p, 1080p and 4k, those three without 4k, or 720p alone; offer 4k only where GET /v1/models lists it); off 720p, re-quote with POST /v1/estimates. Do not send the key on a model whose list has one value: that body is a 400, even when the value is the one it renders.
• language: the language the script is written in. Set it, and show it in the dialogue gate. Write the prompt in that language too — nothing on the API pushes back on a Spanish or Portuguese prompt, and nothing rewrites or judges one either.
• durationSeconds: from the word count, below. Only veo-3.1 defaults to its maximum — seedance-2.0 and seedance-2.5 default to 5, mini to 10, omni-flash to 8, sora-2 to 4 — so always send it. seedance-2.5 defaults to 5 of the 30 it can render, which is the easiest way to pay for the long model and ship a short clip.
• audioEnabled: leave it alone on anything with a spoken line. It exists on the three Seedance variants only — seedance-2.0, seedance-2.5, seedance-2.0-mini — defaults true, and the one time to send it is false, on a clip that is meant to be silent — see below.

Script length → duration

Count the words in the spoken line and round up. Delivery was measured at 2.0 spoken words per second — about 13 characters per second, spaces included, if you would rather count those — with ~0.5s of leading silence before the first word on reference and start-frame renders (2026-08-11, n=5). Plan on 2.0 and give the line air; the older 2.5-to-3 figure was an estimate nobody had timed, and it over-filled every clip it touched.

These tables already pay for that ~0.5s. Whether it is the right budget depends on the render mode — read the scoping note under the first one before you trust it on another.

seedance-2.0 and seedance-2.0-mini — any integer 4 to 15

Script length	Duration
1–8 words	4–5s
9–15 words	6–8s
16–23 words	9–12s
24–29 words	13–15s
30+ words	Too long — offer to split

Leading silence is mode-dependent, and the two things measured here disagree by seconds. Scope the budget to the mode you are actually rendering — do not add both.

What was measured	Leading silence observed	Plan on
Reference or start-frame renders — referenceAssetIds or startImageAssetId	0.0s and 0.44s (n=2), then 0.515s, 0.529s and 0.482s (2026-08-11, n=5 total)	~0.5s. The tables above already pay it
Six seedance-2.0 renders whose mode nobody wrote down (2026-08-02/03)	3.2–3.7s in en; in es, 5.24s, 0.97s and 4.80s on three byte-identical prompts — same body, same duration, three answers	+4s en, +5s es and any unmeasured language, on top of speech

The working rule: budget ~0.5s when you have a reference or a start frame, and check the render. In that mode the table and the arithmetic already agree — 2.0 × (D − 0.5) is exactly the word cap on every rung from 9 words up, and the 1–8 row is deliberately one word tighter than the formula's 9 at 5s — so there is nothing to add. If the result runs past 15s on these two models, the fix is a shorter line, a split, or seedance-2.5 — not a longer clip on a model that does not render one.

When a render does come back with multi-second silence, trim is the remedy, not a bigger budget. Leading silence is trimmable in post; a line clipped by an over-long draw is not. The reserve in the second row is the planning allowance for that mode — reach for it when you have neither a reference nor a start frame, or when you are committing to a duration you cannot re-cut. It costs words: words ÷ 2.0 plus the reservation, rounded into the grid, leaves 22 words at 15s in en and 20 in es, against the table's 29. That gap is the mode, not a rounding error.

How close the reserved case gets is not theoretical. The first es render fit only because rounding up took 12.4s to 13s, and the line finished at 12.88s of 13.07s — a fifth of a second of margin, on a budget that drew near its worst. The silence is drawn fresh every time, in either mode, so the only thing that tells you what you got is the QA step in §7. Check the first silence_end on every render.

If the hook has to land in the first second, drop the eye-contact-break beat from the prompt — that is the beat being paid for. sora-2 measured no leading silence at all on the same prompt, so it is the other way out.

For no-dialogue styles (product hero, premium reveal), default to 15s. The silence budget does not apply: there is no speech to delay.

seedance-2.5 — any integer 4 to 30

Same family, same craft, twice the room. The table above still holds for anything up to 15s; past it, keep planning at 2.0 words per second — the rows below are the same 2.0 × (D − 0.5) arithmetic. Nobody has measured 2.5's leading silence separately, so the mode scoping above is the working rule here too: budget ~0.5s with a reference or start frame, the reserve without one.

Script length	Duration
30–39 words	16–20s
40–49 words	21–25s
50–59 words	26–30s
60+ words	Too long even here — split, or cut the script

A longer clip is not a longer script poured into the same ad. Thirty seconds wants more beats, not slower delivery: a second location, a demo the actor actually performs, a reaction. If the script only fills 15s, render 15s — on seedance-2.0, which is cheaper at that length.

Nobody here has timed a seedance-2.5 render. Do not quote seedance-2.0's 3-to-8-minute fleet range for it; say the wait is unknown and poll.

omni-flash — enum 4, 6, 8, 10

Script length	Duration
1–8 words	4s
9–12 words	6s
13–16 words	8s
17–20 words	10s
21+ words	Too long — split, or move to Seedance

Silence behavior is unmeasured on this model, so this is the 2.0-words-per-second arithmetic and nothing more: the budget is a flat 2.0 × duration, which assumes no leading silence rather than having found none. Verify against your first render — if it front-loads the way Seedance can, the last line is what gets clipped.

sora-2 — enum 4, 8, 12

Script length	Duration
1–8 words	4s
9–16 words	8s
17–24 words	12s
25+ words	Too long — split, or move to Seedance

No silence budget: the one measured render spoke continuously from the first frame. The grid is coarse — there is no 6s and no 10s — so a line that lands between two rungs goes up, never down.

veo-3.1 — enum 4, 6, 8

Script length	Duration
1–8 words	4s
9–12 words	6s
13–16 words	8s
17+ words	Too long — 8s is this model's ceiling. Split, or move to Seedance

Unmeasured here, so these are the 2.0-words-per-second arithmetic and nothing more — and with no leading silence to pay for, the budget is a flat 2.0 × duration rather than the Seedance form. Budget no silence and promise no wait until someone has timed one.

resolution: only where GET /v1/models lists more than one tier, and it moves the price

seedance-2.0 takes 480p, 720p, 1080p, 4k; seedance-2.5 takes 480p and 720p and nothing above. Both default to 720p (2.0 verified live 2026-08-04 against spec 2.6.0; 2.5 read off deployed spec 2.13.0 on 2026-08-07). It is the one output-shape field that is not free — unlike aspectRatio, the tiers are separate credit schedules:

resolution	seedance-2.0	seedance-2.5
480p	≈half the base	≈half the base
720p (default)	base	base
1080p	≈2.5x base	400 — does not exist on this model
4k	≈5x base	400 — does not exist on this model

480p used to cost the same as 720p and no longer does (family reprice, 2026-08-07). The old advice here — "no draft discount, so there is no reason to ask for it" — is retired: it is now roughly half, which makes it the honest tier for a rehearsal render whose only job is to check whether the prompt works.

2.5's missing high tiers are a provider fact, not a rollout gap. Neither provider serves the model above 720p, so it is not coming. Never carry a resolution across a model switch: 1080p that priced clean on seedance-2.0 is a rejected request on seedance-2.5.

These are ratios, not a rate card. Never quote a credit number from this table — it exists so you can warn a user that 4k is a five-fold decision before they ask for it. The number they approve comes from POST /v1/estimates on the exact cell, in this session, like every other price here (gate 2).

Ask for what will actually ship. 720p is right for Reels, TikTok and Stories, where the platform re-encodes anyway; 1080p and 4k are for a client deliverable, a placement with a quality floor, or a render you intend to crop into.

omni-flash depends on the server: GET /v1/models lists 360p, 720p and 1080p for it where the server enables them, and 4k as well where the server publishes that tier (API 2.31.0; never assume it, read the list). Then send resolution, default 720p, and re-quote with POST /v1/estimates off it; Omni's tiers are its own schedule, so never carry the Seedance ratios above over to it. Or it lists 720p alone, and then a body carrying the key is a 400, 720p included. From API 2.34.0 seedance-2.0-mini and veo-3.1 list more than one tier too (mini adds 480p, Veo adds 720p): send resolution there and quote the tier with POST /v1/estimates, never with the Seedance ratios. On veo-3.1 the tier does not move the quote (one flat price), so never offer 720p as a cheaper option; quote it all the same. sora-2 is fixed at 720p and has no resolution field, and sending it the key is a 400. Read the live set from GET /v1/models (resolutions and defaultResolution per model) rather than hardcoding this paragraph.

The seedance-2.0-mini trap, before API 2.34.0 only (from 2.34.0 mini takes the key): POST /v1/estimates accepts resolution: "720p" for mini and prices it (identically to omitting it), but answers 400 invalid_input — "resolution must be one of 720p for seedance-2.0-mini" — for 480p, 1080p or 4k (all four verified live 2026-08-04). POST /v1/videos does not accept the key for mini at all: mini's request variant has no resolution property, and a body carrying one was observed returning 400 Unrecognized key: "resolution" (observed 2026-08-04, not re-verified — re-checking it costs a paid render). So, against a server whose mini resolutions is ["720p"], never send resolution on a mini call. An estimate that passed is not evidence the generate call will.

Output size, measured at 9:16: seedance-2.0 at its 720p default and sora-2 both came back 720x1280 (2026-08-02). omni-flash and veo-3.1 are unmeasured — do not quote a number for them.

Splitting a long script

1. Tell the user the script is too long for one clip and show the word-and-duration math.
2. Offer two options: split at natural sentence boundaries into chunks that each fit, or move to the model with more room (seedance-2.5 tops out at 30s and is the longest single clip here; seedance-2.0 and mini stop at 15). Price both shapes — one 30s call is not automatically cheaper than two 15s ones, and it is a different ad.
3. If they split, each chunk is its own generation call — and the variation count applies to each chunk.
4. Offer to stitch with ffmpeg: download the segments, ffmpeg -f concat -safe 0 -i list.txt -c copy output.mp4, re-encoding if the codecs differ. Hand back both the stitched file and the individual segments.

The full sequence

0. Resolve the product (once per session)

GET /v1/products. The products come back under items, not products (verified live 2026-08-04) — it is the same paginated envelope as GET /v1/generations, so read .items and expect total, limit, offset and hasMore alongside it:

curl -sS https://api.novoads.ai/v1/products \
  -H "Authorization: Bearer $NOVOADS_API_KEY" | jq -r '.items[] | "\(.id)  \(.name)"'

Default to the product named in MASTER_CONTEXT.md under "My workspace". If no default is set: with exactly one product, auto-populate MASTER_CONTEXT.md with its id and name; with several, ask the user once and save the choice. With none, omit productId — it is optional.

Pass productId on every generation call. It is what makes GET /v1/generations?productId=… a useful history later. There is no dated-folder ritual to run here: folders are read-only on this API and there are no projects on it at all.

1. Upload the reference or start frame (skip if there is none)

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}'

Returns assetId, uploadUrl, method, and headers.

Then PUT the raw bytes to uploadUrl, sending exactly the headers it returned:

curl -sS -X PUT "$UPLOAD_URL" \
  -H "Content-Type: image/jpeg" \
  -H "Content-Length: 248193" \
  --data-binary @product.jpg

Both headers are signed into the URL. image/jpeg; charset=utf-8 is a 403, and so is a Content-Length that does not match the bytes you send. sizeBytes must be the real file size, so measure it, do not estimate it.

Upload once, reuse it forever. The assetId is durable and reusable across calls and across models, which is what makes prototyping on seedance-2.0-mini and finalizing on seedance-2.0 cheap. The presigned upload URL expires in 900 seconds; the assetId does not.

Start frame or references, not both. startImageAssetId animates one image as the first frame (on omni-flash it is a reference instead, below). referenceAssetIds (up to 9 on Seedance, up to its own limit on omni-flash, images only; Seedance video and audio references have their own fields, below) composites the images as visual references, addressed positionally in the prompt text as @Image1…@ImageN in the order you send them. They select different modes on the model, so a body carrying both is a 400.

Four models take referenceAssetIds: the three Seedance variants (seedance-2.0, seedance-2.5, seedance-2.0-mini) and, since API 2.30.0, omni-flash. sora-2, veo-3.1 and kling-v3-pro have no such field and their bodies are strict, so sending it is a 400. Every video model takes startImageAssetId, but on omni-flash it is a reference image the model may place anywhere in the clip, not a first frame. omni-flash also takes one referenceVideoAssetId (where the server also publishes firstFrameAssetId, the clip's length bounds durationSeconds) and, only where the server publishes them, firstFrameAssetId (the true first frame), lastFrameAssetId (needs a first frame) and seed; startImageAssetId beside firstFrameAssetId is a 400. Since API 2.35.0 the three Seedance variants also take referenceVideoAssetIds (mp4) and referenceAudioAssetIds (mp3), not combinable with startImageAssetId, capped per kind and in total as the OpenAPI document (GET /v1/openapi.json) publishes, and priced the same as without them; see Seedance video and audio references in the reference. Read which fields a deployment takes from the OpenAPI document it serves, not the version number, and the full rules under omni-flash inputs in reference.md.

2. Price it (gate 2)

POST /v1/estimates with kind, prompt, model, and the duration or image count. See gate 2 above for the exact field list, and remember it rejects any field that does not move the price.

3. Confirm the spoken line (gate 1)

See gate 1 above. Both gates must be satisfied, in either order, before anything is submitted.

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": "...",
    "durationSeconds": 12,
    "aspectRatio": "9:16",
    "language": "en",
    "startImageAssetId": "...",
    "productId": "..."
  }'

Returns 202 with jobId, status, creditsCharged, and model. No warnings — this endpoint does not run the prompt rules at all. If you want them, they came back on the estimate you already made (gate 2); there is no second chance to collect them here.

Set aspectRatio explicitly. Seedance defaults to 16:9 and an ad that ships landscape is a wasted render. Set durationSeconds explicitly too: seedance-2.0 and seedance-2.5 default to 5, mini to 10, omni-flash to 8, sora-2 to 4, veo-3.1 to 8. On seedance-2.5 that default is 5 of a possible 30 — the long model rendering a short clip at the long model's price.

audioEnabled — the one field that only belongs on a silent clip

audioEnabled is a boolean on the three Seedance variants only — seedance-2.0, seedance-2.5, seedance-2.0-mini. It defaults to true, so omitting it renders exactly what this endpoint rendered before the field existed. omni-flash, sora-2 and veo-3.1 have no such property and their bodies are strict, so sending it to one of them is a 400, not a field quietly dropped on the way to a paid render.

Send audioEnabled: false when the clip is meant to be silent — a pixar or claymation beat that gets its voice-over laid in post, a product cutaway built to run muted. Otherwise the model generates a voice track and sound effects that get thrown away, or worse, an invented narrator over a film that was supposed to be wordless.

It does not change the price, which is why POST /v1/estimates does not take the field — sending it there is a 400 on an otherwise valid estimate body.

Keep the silence sentence in the prompt as well. The prose clause (a silent product film with no spoken dialogue) and the flag are belt and suspenders and do different jobs: the flag mutes the render, the clause stops the model from staging a talking shot in the first place — an actor mouthing nothing, billed in full. Dropping the clause because you set the flag trades a composition fix for a mute button.

For N variations, fire the identical payload N times. Five generations per organization may be in flight at once — fire at most five, then start the next as each one reaches a terminal state. A sixth submission comes back 429 with error.details.reason of concurrency_limit, which is a different problem from a rate limit and takes a different response: wait for a slot, do not lengthen the backoff.

Log each submission immediately: one line appended to logs/novoads-api.jsonl with the timestamp, endpoint, model, jobId, productId, creditsCharged from this 202, and the request config (duration, aspectRatio, language, reference counts, prompt word count). The charge goes on the line now — the poll payload does not carry it, so a line written without it can never be completed. Never log the prompt text, the key, or the Authorization header. The log is observability — latency and spend after the fact. It is never a pricing source. Prices come from /v1/estimates, always. The file is gitignored: it is the user's session history, not repo content.

5. Poll

curl -sS https://api.novoads.ai/v1/generations/$JOB_ID \
  -H "Authorization: Bearer $NOVOADS_API_KEY"

status is one of queued, running, finalizing, succeeded, failed, blocked, canceled. The last four are terminal. queued means charged and submitted but not yet rendering, which is normal and not a stall.

Poll until TERMINAL, not until succeeded. A loop that waits only for succeeded never returns on a job that failed, and the user watches a spinner forever on a render that is already dead.

Poll every 15 seconds. Not 5: five concurrent jobs at a 5-second interval is 60 calls a minute, exactly the per-key rate limit, with no headroom left for the calls that do real work.

Tell the user the wait up front, per model, so they do not think it hung — and quote it as a range, never as a promise:

• seedance-2.0: fleet range 3 to 8 minutes, median around 5. Observed here: ~171s, ~171s, ~154s on three renders (2026-08-02/03).
• seedance-2.0-mini: fleet range 2 to 3 minutes.
• seedance-2.5: nothing published and nothing measured. Say unknown — do not lend it seedance-2.0's range on the strength of the family name.
• sora-2: no fleet range published. Observed here: ~123s (2026-08-02, n=1).
• omni-flash, veo-3.1: nothing measured and nothing published. Say the wait is unknown rather than borrowing Seedance's.

Do not promise five minutes. The two renders anyone here has actually timed both came back under three, and a user told "about five" who gets it in two is fine, while a user told "about five" who waits nine has been misled. Give the range, name the model, and say the numbers are ranges.

When status is succeeded, outputUrl is a presigned download URL valid for outputUrlExpiresInSeconds (3600). Read the job again for a fresh one rather than storing it. Update that job's log line with the terminal status and the elapsed time.

creditsCharged is NOT on this payload — do not go looking for it here. The poll returns exactly createdAt, error, jobId, kind, model, outputUrl, outputUrlExpiresInSeconds, prompt, status (verified live 2026-08-03). The charge is on the 202 from the submit, and that is the only place it is ever available. Carry it forward from the line you already wrote; if it is missing there it is gone, and you never reconstruct it from a rate. The keyed, portable line-update recipe is in logs/README.md at the repo root — update by jobId, not by position, because up to five jobs are in flight and the line you want is often not the last one.

6. Download and hand it over

OUT_DIR="outputs/seedance-ugc-cerave"      # descriptive, not the job id
mkdir -p "$OUT_DIR"                        # outputs/ is gitignored and absent on a fresh clone
curl -sSL -o "$OUT_DIR/ad.mp4" https://api.novoads.ai/v1/generations/$JOB_ID/watch \
  -H "Authorization: Bearer $NOVOADS_API_KEY"

The mkdir -p is not boilerplate. outputs/ is gitignored, so it does not exist in a fresh clone, and curl's failure when the directory is missing is curl: (56) Failure writing output to destination, passed 559 returned 4294967295 — which reads like a broken download of a render that actually succeeded and was already billed. Create the directory first and that whole detour disappears.

/watch 302s to a URL signed at request time, so it never hands you an expired link. While the job is unfinished it is a 409 naming the current status.

Save under outputs/<descriptive-subfolder>/ — outputs/seedance-ugc-cerave/, not the job id — and name the file for what it is. Then always open the folder so the user can review immediately: open "<dir>" on macOS, xdg-open on Linux, explorer on Windows. Try open first and fall back silently.

Present multiple variations as a numbered list. If a job came back failed or blocked, say which, and quote the error the job carries.

7. Video QA (mandatory)

Run this on every video before you hand it over. It costs no credits and takes seconds. A video can be technically succeeded and still be undeliverable in ways that no frame and no waveform will show you: the brand name spoken as a different word, a third of the clip spent in silence before the hook, captions burned in, a silent track on an ad that was supposed to talk. Watching it back is not enough either — you will hear what you expect to hear, because you wrote the line.

The failure this catches is real and it is the expensive one. On 2026-08-02 a seedance-2.0 render spoke the approved 17-word line verbatim and pronounced "Novoads" as "Nuvenov's". Every frame looked perfect. Only the transcript caught it.

1. Container, stream and duration

ffprobe -v error -show_entries format=duration -show_entries stream=codec_type,width,height \
  -of default=noprint_wrappers=1 ad.mp4

Check: the duration is the one you paid for, width/height match the aspectRatio you sent (9:16 comes back 720x1280), and there is an codec_type=audio stream at all. No audio stream on a clip with dialogue means the render is dead — that is not a QA note, that is a re-render.

2. Levels and silence

ffmpeg -i ad.mp4 -map 0:a -af volumedetect -f null -
ffmpeg -i ad.mp4 -map 0:a -af silencedetect=noise=-35dB:d=0.3 -f null -

Do NOT add -v error to these two. Both filters print their findings at ffmpeg's info level, so -v error suppresses the entire result and the command exits 0 having told you nothing. It looks exactly like a clean pass. This is the single easiest way to run video QA and learn nothing from it — verified 2026-08-02: with -v error the silencedetect call printed no output at all on a clip carrying 3.7s of leading silence.

ffprobe in step 1 is the opposite case: -v error there is correct, because -show_entries writes to stdout.

Check mean_volume is not near-silent (the renders measured here came back −19 to −25 dB), and read the first silence_end — that is when the first word actually lands. Compare it against the silence budget in Script length → duration: on Seedance expect 3–4s in en, and anywhere from 1s to 5s in es, drawn fresh each render. Much more than that on a short ad means the hook is gone.

3. Transcribe — the only check that hears the brand name

Doctrine: shared/references/craft.md § 1. The commands below are this surface's; the rule is not this skill's.

ffmpeg -y -v error -i ad.mp4 -ar 16000 -ac 1 -c:a pcm_s16le /tmp/ad-qa.wav

Then transcribe /tmp/ad-qa.wav with whichever Whisper the machine has — the same one analyze-video and clone-video-ad use:

whisper /tmp/ad-qa.wav --model base --output_format txt --output_dir /tmp   # openai-whisper
whisper-cli -m <path/to/ggml-medium.bin> -f /tmp/ad-qa.wav -l en --no-prints  # whisper-cpp

Read the transcript against the line the user approved at gate 1, and check three things:

• The brand name came out as the brand name. This is the whole reason the step exists. If it did not, the fix is the phonetic spelling in gate 1 — not a re-roll of the same prompt.
• The words are the approved words. A dropped or invented clause means the render does not match what was signed off.
• language matches what was rendered.

What to do when QA fails

Say which check failed and show the evidence — the transcript line, the silence figure, the missing stream. Then:

• A mispronounced invented brand name → re-render with the phonetic spelling for that language (gate 1): NO-vo-ads in en, Novo Ads in es. A prompt fix with a validated result on seedance-2.0, not a gamble.
• A stutter — a clause spoken twice. Seen once in three es renders of the same payload. Re-render the identical body: the defect is nondeterministic delivery, not a prompt fault, so re-sampling the same cell is the right move. This is the one exception to "never resend an identical payload", which is about prompt-caused image defects. It still costs credits, so it still goes through gate 2.
• Leading silence eating the hook → re-render shorter, or drop the eye-contact-break beat, or move to sora-2, which measured none.
• No audio stream, or a silent track on a talking ad → re-render. Nothing in the prompt fixes a dead track.
• Burned-in captions → the clean-plate clause is missing from the prompt. Add it and re-render.

Every re-render is a new charge, and unlike image QA there is no automatic-retry allowance here: a video re-render goes back through gate 2. Show the QA finding, show a fresh estimate, and get a yes.

Burned-in captions: POST /v1/captions

This API has a first-party captioning endpoint (verified live 2026-08-04 against spec 2.6.0). It was undocumented in this pack, and the decision tree used to send every caption request to the local ffmpeg skill as though no endpoint existed. Both paths are real — see Which caption path below.

What it does: takes a video, transcribes its own audio, burns styled subtitles into it, and gives you back a new MP4. Asynchronous, exactly like POST /v1/videos.

What it does not do: THIS endpoint never returns caption text, timings, or an SRT. There is no transcript field anywhere on a caption job — GET /v1/generations/{jobId} carries outputUrl and nothing else about the words. That is still true and is not a bug to file.

But the API as a whole does return the words now. POST /v1/transcripts gives you text, word-level timings and an SRT in one synchronous call, on the base video, with no local install — see POST /v1/transcripts in reference.md. So "the user wants an SRT" is no longer a reason to leave the API; wanting to edit the wording before rendering, or wanting a style outside the 30 presets, still is.

There is nothing to write. The text is transcribed from the audio. Your only choice is preset.

The call

curl -sS -X POST https://api.novoads.ai/v1/captions \
  -H "Authorization: Bearer $NOVOADS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jobId":"<a succeeded video job>","preset":"casper"}'

Field	Notes
jobId	A video this API generated, via POST /v1/videos.
assetId	A video you uploaded, via POST /v1/uploads. Must be a video, not a still.
preset	Required, no default. One of 30 styles — the tiers are priced differently, so nothing could be a safe default.

Exactly one of jobId and assetId. Sending both is a 400, not a guess — captioning the wrong one of two sources still bills you.

POST /v1/videos/{jobId}/captions is the same operation with the source in the path. It is the natural call when the source is a job, and it cannot express the upload case at all (an assetId contains slashes and is not a path segment). Either one is fine; prefer POST /v1/captions if you want one code path for both sources.

Response is 202 with jobId, status, creditsCharged, and model (always captions-v1 — this endpoint has one renderer and it is not in GET /v1/models, because a caption is applied to a video rather than generating one). Poll the returned jobId at GET /v1/generations/{jobId} to a terminal status, then …/watch for the file — the same sequence as a render, and the caption job is a separate row with its own status and its own refund.

Presets and price

GET /v1/caption-presets lists all 30 styles, each carrying its own tier and its own rate. There are two tiers: basic, and the costlier dynamic (context-aware and animated). Read the rate off that endpoint and confirm it with the caption arm of POST /v1/estimates — a rate written into this file is a rate that rots, and the one you quote the user must be the one the API just gave you. Its own endpoint rather than a kind on GET /v1/models, because a caption style generates nothing.

Billing is per minute of source, rounded up, one-minute minimum — so everything this API generates (≤30s, on seedance-2.5; ≤15s everywhere else) bills exactly one minute at the tier rate. Above the 1080p tier it doubles again, measured on the short edge: an ordinary portrait 1080x1920 is 1080p held sideways and is not doubled; a true 4K source is, and an omni-flash render at 4k (where GET /v1/models lists it) is one, so captioning it bills doubled. Quote it from the caption arm of the estimate, never from this line.

Duration and resolution are read from the file itself at request time, not from anything you declare. Gate 2 still applies — price it with the caption arm of the estimate, which charges nothing:

curl -sS -X POST https://api.novoads.ai/v1/estimates \
  -H "Authorization: Bearer $NOVOADS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"kind":"caption","preset":"casper","jobId":"<source job>"}'

Name the source in the estimate for anything longer than a minute. A quote without jobId/assetId is the one-minute minimum, and a 10-minute upload costs ten times it. (Verified live 2026-08-04: sourceless casper quoted 0.4, sourceless glass quoted 0.8. As always, the number you show the user is the one that came back, not one from this page.) A caption estimate has no prompt, so it returns no warnings.

What bites

• 409 — the source job has not succeeded yet, or it was rendered with audioEnabled: false. No speech, nothing to transcribe, and the API refuses rather than charging you for an empty result. Silent b-roll cannot be captioned.
• 404 — a jobId from the dashboard rather than the API answers 404, identically to one that does not exist. So does an asset outside your organization or an incomplete upload.
• 400 — a still passed as assetId, both source fields at once, or a file that cannot be measured. A source we cannot measure charges nothing, because a file we cannot price is one we cannot bill.
• Re-captioning the same video in the same style is safe and free. The second call returns the first job's id and charges nothing. A different preset on the same video is a new job and a new charge — so iterating on style costs money each time; pick with the user before you call.
• 429 with details.reason: caption_concurrency_limit — 10 concurrent caption jobs, counted separately from the 5-generation render ceiling, so a batch of captions can never block your next render.

Which caption path — the API or the local skill

Both exist. Present both and let the user pick; do not silently default.

	POST /v1/captions	the caption-video skill
Cost	Costs credits (gate 2 applies)	Free
Setup	None — one API call	Heavy: Whisper, HyperFrames, a working Node/npm project, and an ffmpeg alpha composite. Homebrew ffmpeg ships without libass, which is the trap the local guide exists to route around
Styles	30 presets, fixed	Anything you can write in HTML/CSS/GSAP
Output	A new MP4, subtitles burned in	A new MP4, subtitles burned in
Transcript / SRT	Not from this endpoint — but POST /v1/transcripts returns text, word timings and an SRT, billed per minute of source and priced through POST /v1/estimates, no install	Yes — Whisper gives word-level timings you keep and can re-use
Control over wording	None — transcribed, not authored	Full — edit the transcript before rendering
Source	An API jobId, or any video you upload	Any local file

Rules of thumb. One finished clip, a standard look, no local toolchain → the API. A caption style outside the 30, hand-corrected wording (invented brand names are the usual reason), a needed SRT, or a batch big enough that per-minute credits add up → the local skill. If the video was rendered with audioEnabled: false, the API cannot caption it at all.

Images come back in the response, or answer running

POST /v1/images usually returns the finished images in the response body (status: "succeeded"). Since API 2.39.0 a render still going after about 105 seconds answers 200 with status: "running", the same jobId and images: []. That is a paid render in progress, not a failure: never resubmit it. Poll GET /v1/generations/{jobId} every 15 seconds to a terminal status; on succeeded the job carries images[] in order, the same entries the response would have carried, and on failed or blocked the credits are refunded. Log creditsCharged from the first response. There is no /watch step for images. The response carries images[] with url, expiresInSeconds, assetId, width, and height, plus jobId, status, model, and creditsCharged. Chain downstream calls from assetId, not from url — the URL is a one-hour presign and nothing downstream accepts it. No warnings here either. The prompt rules run on POST /v1/estimates only — and an image estimate does return them, from rules written for images: banned_polish and blank_label observed live on deployed spec 2.19.0 (verified 2026-08-12). The older claim here, that an image estimate came back with no warnings key at all, described the 2026-08-04 deployment. Read and judge them like the video ones; they are advisory and cannot refuse or reprice anything, so the image prompt libraries are still the real check.

• referenceAssetIds: images only, order preserved and addressable positionally by the prompt. Upload each one first — there is no base64 field. The cap is per model, not one number: nano-banana-pro takes up to 14, reve-2.1 up to 8, gpt-image-2 up to 4, and seedream-5-lite and seedream-5-pro (API 2.39.0) up to 10; Seedream has no quality, resolution or outputFormat, and its aspect ratios come from GET /v1/models. The bodies are strict, so a fifth reference to gpt-image-2 is a 400, not a silently dropped image — which is the good outcome, because a dropped reference is a paid render missing the product.
• numImages: 1 to 4, and it multiplies the price. quality (high, medium, low; default medium) and outputFormat (png, jpeg, webp; default png) on the GPT Image models, resolution (1K, 2K, 4K; default 2K) on nano-banana-pro (API 2.36.0). quality and resolution move the price: quote the exact cell with POST /v1/estimates using the same fields. Read the model's lists from GET /v1/models; image resolution is matched exactly, so 4k is a 400.
• Images take no startImageAssetId: there is no first-frame concept on a still. A preset takes its one photo as imageAssetId instead (API 2.40.0, decision tree above).

Image QA (mandatory)

Images are money already spent, and a defective still is not a deliverable. After each image comes back, look at it.

Look for: extra or missing hands or fingers, wrong limb count, distorted or duplicated or merged facial features, melted or fused objects, impossible anatomy, stray limbs, texture and boundary artifacts, and garbled text where text was requested.

If something is wrong, regenerate with a corrected prompt that names the defect ("exactly two hands, five fingers each", "a single face, no duplicated features"). Never resend the identical payload expecting a different result.

• Cap: 2 regenerations per originally requested image, 3 attempts total. After the cap, stop, show the best attempt, say what still looks wrong, and ask how they want to proceed.
• No second cost confirmation. Once the batch is approved, QA retries proceed automatically — the one exception to gate 2.
• Every retry is billed. Sum the extra credits and report them when the loop ends.

The references/ folder

Check it before asking the user for anything. references/influencers/ for people, references/products/ for products, references/aesthetics/ for style and mood. If a relevant file is already there, offer to use it instead of asking.

It ships empty. Do not stall on it.

references/ is gitignored, so on a fresh clone it holds nothing but empty folders. That is the expected state, and "go find me a photo" is the wrong first move for most requests. Ask what they are advertising first, then decide whether a photo is needed at all:

What they are advertising	Does it need a reference?
Software, an app, a SaaS product, a service, a course, an agency	No. There is no physical object to hold. The actor talks to camera and the product never appears — this is the single most common ad shape and it needs nothing in references/. Route it and go.
A physical product the user sells	Yes, and it is worth waiting for. Ask for one photo, put it in references/products/, pass it as startImageAssetId or in referenceAssetIds. Do not invent packaging: the model will render a plausible fake label and charge you for it.
A specific person's likeness	Yes. references/influencers/. Without it Seedance re-casts on every render.
A look, a palette, a lighting mood	Optional. A style board sharpens it; prose can carry it.

Never substitute a blank or generic product for one the user has not given you. If they are advertising something physical and have no photo to hand, say what you cannot do rather than rendering a nameless bottle. Nothing on the API will warn you, and the charge lands either way.

Before uploading a reference, if its longest side is under 1024 px, upscale with Lanczos to 1080 px on the long side and re-encode as RGB JPEG at quality 90–95, which strips alpha and keeps the payload sane. (No minimum input size is documented for this API — the practice carries over from a sibling API that answered small images with a 422, and is unverified here.)

Errors: branch on error.code, never on the message

Every error is {"error":{"code":..., "message":..., "requestId":..., "details":...}}. Quote requestId when reporting a problem; it matches the x-request-id response header.

Status	code	What it means	What to do
400	invalid_input	The request is malformed — an unknown key, an out-of-grid durationSeconds, a prompt over the model's ceiling. details.issues names each bad field. Never a judgement on the writing.	Fix the field. Nothing was charged.
401	unauthorized	Missing, malformed, or revoked key.	Send the user to https://novoads.ai/dashboard/settings?tab=api to create a new one.
402	insufficient_credits	details carries required and available.	Tell the user the gap. Do not retry.
403	forbidden	details.reason is plan_required, subscription_inactive, or the API is off for that account.	Say which. These are different fixes.
404	not_found	No such object for this organization.	Do not assume it exists elsewhere.
409	conflict	Includes /watch on an unfinished job.	Keep polling.
422	content_policy	Moderation blocked it, and this is the only way a prompt is refused for what it says. The estimate skips moderation, so it can land on a prompt the estimate priced clean.	Nothing was charged. Rewrite or stop.
429	rate_limited	Four different causes. Branch on details.reason.	See below.
500	internal_error		Do not blindly retry. See below.
502	provider_failed	A model provider failed.	Credits are refunded automatically.

The 400 vs 422 line is simple now, and worth stating because it used to be blurred. A 400 is a malformed request and nothing else. A 422 is moderation and nothing else. Prompt craft — a missing actor descriptor, no spoken line, a forbidden word — produces no status at all: the API neither refuses it nor mentions it. Do not write a retry loop that expects a rule to stop a bad prompt; the only thing that stops it is you, before you send it.

A 429 is not always a rate limit. details.reason names which of four ceilings refused you, and they are paced differently:

• concurrency_limit — five generations already in flight for the organization (details.inFlight says how many). Waiting is the only fix: a longer backoff does nothing, a finished job does. Generation endpoints only; reads, estimates, and uploads never hit it.
• key_limit — 60 requests a minute on this key. The X-RateLimit-* headers track this one. Honor Retry-After.
• organization_limit — 180 a minute across every key the organization holds. X-RateLimit-* will still show room on your key; that is correct, not a broken limiter. Minting another key does not raise it.
• client_limit — 1,200 a minute pre-authentication. Carries no X-RateLimit-* trio.

Every 429 carries details.reason and a Retry-After header — sleep on the header, branch on the reason. details.inFlight (concurrency refusals) is the only other key the spec names; do not expect one it does not.

On a 500, do not retry until you have checked. The generation endpoints charge credits, and a failure can land after the debit committed, so a blind retry can pay twice. Call GET /v1/generations first. If the job is there, poll it instead of resubmitting. There are no idempotency keys.

Guardrails

Append one line per new failure. Forward only. Every bullet is a real thing that went wrong.

• Never write a credit number into a file, a summary, or MASTER_CONTEXT.md. Prices come from /v1/estimates at call time.
• Never generate without both gates: the spoken line, and the cost.
• Always set aspectRatio. The Seedance default is 16:9.
• Always set durationSeconds. seedance-2.0 and seedance-2.5 default to 5, mini to 10, omni-flash to 8.
• Only seedance-2.5 renders past 15s (4–30). Asking any other model for 20 is a 400, never a rounded-down render.
• seedance-2.5 has no 1080p and no 4k. Do not carry a resolution across a model switch.
• Content-Type on the presigned PUT must match byte for byte. Adding ; charset=utf-8 is a 403.
• Never resubmit after a 500 without checking GET /v1/generations first.
• A 400 is a malformed request. Read details.issues, fix the field, and do not go looking for a prompt rule — none of them can 400 anymore. Never send styleFamily: the field no longer exists on this API and any body carrying it is a 400.
• Prompt rules exist only on POST /v1/estimates, as the advisory warnings array, and they cannot refuse or reprice anything. Nothing on the API will stop a weak prompt from being rendered and billed, so the prompt libraries are still the quality gate. Read every warning, judge it against your prompt (they false-positive — see gate 2), and never apply a suggested fix without checking it fits.
• Never fire more than five generations at once, and poll at 15 seconds, not 5. A QA retry still costs credits: cap at 2, and report the extras.
• Never hand over a video you have not QA'd. §7 is free and it is the only thing that hears the brand name. A render that looks perfect frame by frame can still say the wrong word.
• Never put -v error on the volumedetect or silencedetect calls. It suppresses the results and the check reports nothing while looking like it passed.
• Spell invented brand names phonetically inside the quoted line, and re-run gate 1 when you do — the words changed.
• audioEnabled exists on the three Seedance variants and kling-v3-pro. Sending it to omni-flash, sora-2 or veo-3.1 is a 400, and sending it to /v1/estimates is a 400 on any model.
• Image referenceAssetIds caps are per model: 14 on nano-banana-pro, 8 on reve-2.1, 4 on gpt-image-2, 10 on each Seedream 5 model. There is no single number.
• Real brands only in prompts. Do not substitute a blank bottle for a product the user has not given you — the API will happily render it and charge for it, and nothing will tell you. Ask for the photo.

References

• reference.md — every endpoint, field, limit, and error, plus the dated discrepancy list.
• prompting/prompt-library/seedance-2-ugc.md — the UGC prompt formula, with a worked example that passes validation.
• prompting/guide.md — marketing brief → API.
• prompting/brand-voice-starter.md — template to copy into MASTER_CONTEXT.md.
• The other Seedance formulas: seedance-2.md (platform guide), premium reveal, product hero, studio lookbook, feature walkthrough.
• The other video models: sora-2.md, veo-3-1.md. kling-3.md is vendor prompt craft for kling-v3-pro; the request body comes from reference.md.
• Image and character craft: ugc-selfie-style.md, ugc-product-selfie.md, product-showcase.md, influencer-recreation.md, character-sheet.md, character-sheet-gpt-image-2.md, nano-banana.md.
• https://api.novoads.ai/v1/openapi.json — the spec itself, which is the authority when this file and it disagree.