Ugc Ads

/pika:ugc-ads

Parameters

Cost transparency gate

Before any paid MCP call, call identity_balance({verbose: true}) once. Surface the current balance, recent burn rate, and remaining runway, then gate the run with this exact message:

Estimated cost: about 4,000 credits (~$40) for a typical Seedance UGC ad with fallback/retry budget. This exceeds $5, so Reply proceed to continue or cancel to stop.

Do not call any paid MCP tool until the user replies proceed. If the user replies cancel, stop without generating. The gate runs after the product URL is known and before avatar analysis, screenshot capture, video generation, captions, or paid retries.

Runtime expectations

Typical end-to-end run: 6–12 minutes. Breakdown:

• Step 1 (WebFetch) + Step 3 (capture_website screenshot): ~10–30s
• Step 7 (generate_reference_video): ~3–5 min for seedance, ~5–7 min for kling
• Step 7b/c (cartoonize + retry): adds ~1–2 min if seedance moderation rejects the avatar
• Step 8 (deterministic brand/spec overlays): 1-2 edit_text_overlay calls, ~30s–5 min total
• Step 9 (captions): single add_captions call, ~30s–5 min (transcribe + burn in one shot)

If the run exceeds 15 min without progress, something is wrong — inspect the tool-reported generation status and error message.

Pre-generation wall-clock guard

Start a timer at skill start once the product URL is available and the cost gate has passed. Time spent waiting for the user's proceed reply is not prep time and must not trigger this guard. The first paid generation call is generate_reference_video, the long-pole paid stage, and it must be invoked within 5 minutes of skill start. If you have not invoked generate_reference_video within 5 minutes of skill start, stop before any paid generation call and report failed_pre_generation_timeout with what you have so far: fetched product facts, chosen category, avatar source, screenshot status, draft dialogue, and the exact blocker. Do not keep refining script wording, prompt grounding, or shot order.

Print a single-line progress checkpoint after each prep stage and right before the paid generation call:

• Stage 1/3 done — product fetched and categorized.
• Stage 2/3 done — avatar and screenshot ready, composing dialogue.
• Stage 3/3 done — prompt locked, calling Seedance now.

Script and prompt iteration is maximum 2 passes. After the max 2 passes, ship what you have to generate_reference_video; do not continue polishing the hook, punchline, or screen-close-up wording.

Long-running task_status polling

When any long-running generation or edit call returns a task_id with or without an initial status, including {task_id}, {task_id, status: "queued"}, or an initial queued, running, or processing status, record the task id and start time immediately.

• Call task_status({task_id}) in a tight loop until terminal (completed | failed | cancelled). No manual sleep and no Bash polling; the worker holds each status call open.
• Emit ONE visible progress line every 60s while status is queued, running, or processing: Seedance i2v queued for {N}m {S}s... still processing. Replace the provider/stage label when polling Kling, GPT-image-2, caption, or edit tasks.
• On completed, unwrap the returned result URL and continue.
• On failed or cancelled, surface failure to the user with task_id, status, and the last status message.
• After 15 min total from the original submit, call task_cancel({task_id}) if the task is still non-terminal, then surface failure to the user. If cancel reports the task is already terminal, call status once more and report that terminal result.
• Do not submit a duplicate request while the original task is still queued, running, or processing.
• Async polling budget: one active task with one 15 min polling window. After the polling cap is exhausted, cancel the task and surface failure instead of submitting another paid render unless a later step explicitly allows a separate capped retry after the original task is terminal.

Engine choice: Seedance default, Kling fallback

Default to Seedance for UGC selfie/talking-head ads because it handles native lip-sync, single-prompt multi-beat pacing, and optional 3:4 output well. Use Kling when the caller explicitly passes provider=kling, or after Seedance exhausts the capped cartoonized retry only if the user chooses Kling from the stop message. Kling's tradeoff is stricter aspect-ratio support but a separate moderation path and explicit shot segmentation.

Steps

0. Resolve input (empty-args menu)

Strip flags and key=value parameters from $ARGUMENTS. If no product URL remains and there is no usable product URL in prior context, print this menu and stop:

Which product should the UGC ad promote? Required:

• Product URL — page to fetch for product name, category, visual references, and language

Optional: avatar_url=, provider=seedance|kling, aspect_ratio=9:16|3:4, variants=9:16,16:9,1:1, category=auto|HAUL|APP|FOOD|BEAUTY|FITNESS|TECH, captions=true|false.

If the product URL is present, skip this step silently.

1. Fetch + categorize

WebFetch the URL: pull product_name, brand_name, value prop, brand color, product form, packaging, hero copy, target user, category, and the primary language of the page. Use category= if passed; else trust the WebFetch signal; fall back to HAUL for physical, APP for digital.

Build two grounded fact lists from the fetched page:

{
  "grounded_specs": [
    {
      "claim_text": "250W total output",
      "value": "250",
      "unit": "W",
      "source_quote": "visible source-page text containing the exact value",
      "source_url": "<product URL>"
    }
  ],
  "claims_allowlist": ["250W total output"]
}

Every numeric spec claim (W, mAh, %, minutes, ports, price, dimensions, counts, charging speeds, battery size, rankings) must come from visible source-page text and include a source_quote. If the source page does not visibly support a number, leave it out of grounded_specs and claims_allowlist. Do not infer specs from product category, model name, common knowledge, or competitor pages.

2. Resolve the avatar (fallback to built-in if missing)

• If avatar_url was passed → use it as-is.
• If NOT passed → use this built-in fallback:

  https://cdn.pika.art/v2/files/agent/17d62bf9-0edb-49e4-9ba9-2c5419fa518f/seedream-1777624057811.jpeg
  

  Pre-generated 3D animated Pixar-style portrait of a young female creator — pre-cartoonized so seedance moderation accepts it directly, neutral enough to fit any category. Note in the final summary that the fallback was used so the caller knows to supply their own portrait for persona consistency next time.

2.5 Avatar-type probe for creator portraits

Run this probe before Step 7 and before any paid generate_reference_video call. It applies to caller-supplied avatar_url, the built-in fallback, or any creator portrait chosen as @Image1. The built-in fallback is already a non-IP stylized creator, but still document its source in the final summary.

Call analyze_media once:

query: "Classify this image for paid video generation. Is it a photograph of a real human face, an AI-generated realistic portrait, a stylized / illustrated character, or a recognizable trademarked / copyrighted character such as Batman, Pikachu, or Mickey Mouse? Return strict JSON only: { \"avatar_type\": \"real_human\" | \"ai_realistic\" | \"stylized_illustrated\" | \"recognized_ip\", \"recognized_character\": string | null, \"moderation_risk\": \"low\" | \"medium\" | \"high\", \"recommendation\": \"proceed\" | \"warn\" | \"reject\" }. Use null for `recognized_character` when no specific character is recognized; never write \"none\", \"unknown\", or explanatory prose in that field."

Route from the result:

• recognized IP / copyright risk -> STOP only when avatar_type is "recognized_ip", or recognized_character names a specific character (for example "Batman"), or when both moderation_risk is "high" and recommendation is "reject". Treat recognized_character: null, empty string, "none", "unknown", "n/a", and low/medium moderation_risk as not enough to stop by themselves. Run this check before the real/stylized routes. A chibi Batman is still Batman even when avatar_type is stylized / illustrated.
• real human / AI-generated realistic -> proceed normally.
• stylized / illustrated -> proceed with a visible warning that stylized avatars may be accepted by Kling but can be inconsistent under Seedance moderation; continue only if the user supplied or accepted that avatar.
• trademarked / copyrighted -> STOP before generation. Surface this message: The avatar appears to be a trademarked character ([X]). Most video providers will moderate this and refuse to generate. Pass avatar_url=<real-looking-photo-url> to override.

3. Capture the product screenshot (best-effort)

Call capture_website with mode: "screenshot". Use mobile=true for handheld-product categories (APP / FITNESS / BEAUTY) so the captured page renders as a portrait phone screen; mobile=false for desktop-context categories (HAUL / TECH / FOOD).

If the call fails (timeout, browser pool down), retry once. If still failing, proceed without the screenshot — the skill is degraded but functional. The close-up beat then describes the page from prose only and Beat 2's reference_images is just [avatar_url].

Capture URL → screenshot_url (or null).

4. Compose the prompt

The full prompt is a single multi-beat string passed to one generate_reference_video call. Structural prose (not markdown bullets). Every beat has a Says: "..." line for lip-sync. Pacing target ~5.5–6 words per second across the whole 15-second ad (≈85–90 words total). @Image1 is the avatar, @Image2 is the screenshot when available.

Write all Says: "..." lines in the language detected from step 1's WebFetch. Both seedance and kling lip-sync handle multilingual; if the product page is Chinese / Japanese / Spanish / etc., the dialogue should be in that language. Hook archetypes from step 5 are language-agnostic — adapt the rhetorical move to the language's natural register.

Spec-grounding rule: forbid inventing numbers. Any spoken or visual number/unit claim must appear verbatim in claims_allowlist. If a number is not in claims_allowlist, rewrite the line qualitatively ("charges fast", "multiple ports", "big battery") or omit the claim. Do not say "50% in 28 minutes", "3 ports", "140W", prices, counts, or time windows unless that exact claim is source-backed.

Brand/spec text rendering rule: Do not ask Seedance or the video model to render the brand wordmark, product wordmark, packaging label, or spec text from prose. Video-model text comes out garbled. The prompt may show @Image2 as a reference, but any new brand name or spec copy that must be readable is added later by the deterministic Step 8 overlay.

HOOK (0–3 sec) <visual setting + creator framing + face/body cue>. Says to camera, fast and energetic: "<hook line>". <style anchor — POV handheld, authentic, raw>.

JUMP CUT 1 (3–6 sec) <wide POV — creator's body language, product partially in frame edge>. <face cue>, says fast: "<setup line>".

JUMP CUT 2 (6–9 sec) <next visual beat — could be the screen close-up showing @Image2 OR another reaction beat, depending on which beat the dialogue arc puts the reveal>. Says (or voice continues over the shot if it's a screen close-up), fast and confident: "<reveal line>".

JUMP CUT 3 (9–12 sec) <next visual beat — same logic; one of the JUMP CUTs is the screen close-up, the others are wide-POV reaction shots>. Says, fast: "<insight twist line>".

OUTRO (12–15 sec) <selfie POV, mid-chest framing, same setting>. Says to camera, fast: "<punchline line>".

avatar is image 1, asset is image 2

Screen-close-up beat — exactly one across the ad, position is dialogue-driven:

• Place the screen close-up on whichever JUMP CUT (1, 2, or 3) the reveal line lands on. Most ads put it on JUMP CUT 2; if the narrative needs it earlier or later, JUMP CUT 1 or JUMP CUT 3 is fine. Pick by content, not by slot number.
• The screen close-up beat shows @Image2 exactly as-is and includes ONE finger-point gesture (a single finger entering from the frame edge, pointing at the hero text or product — no tap, no swipe, no scroll, no hover-on-CTA). The point gesture is the only screen interaction in the entire ad.
• The other JUMP CUTs are wide-POV reaction beats: hands stay on knees, on the bed, or at sides.

Trust @Image2 — when the product page is shown, reference the image; do NOT describe its UI in prose. Describing UI triggers the model to invent extra panels / dropdowns / sidebars / animations. Reference the image; trust it.

5. Category essences

Each essence is the brief you read before composing the 5 beats. Pick one from category in step 1 and write the actual Says: "..." lines tailored to the real product.

Category numeric guard: category archetypes are rhetoric/camera guidance, not permission to invent dates, durations, counts, discounts, costs, ratings, charging times, port counts, output figures, or urgency windows. Any numeric phrase must be rewritten unless the exact claim appears in claims_allowlist. When claims_allowlist lacks a matching number, use non-numeric language such as "after using it", "during the routine", "the source-backed spec", "launch offer", or "available now".

HAUL_UNBOX

• When to use & why: fashion, handbags, jewelry, shoes, designer drops, streetwear, luxury cosmetics with packaging story, accessories — anything where brand packaging + texture/material is the value prop. Viewers convert on vicarious-unboxing dopamine + "I just got this" social proof; texture and hardware ARE what the customer pays for, so the close-up lands on materials, not function. Not TECH (→ TECH_UNBOX), not skincare/makeup application (→ BEAUTY_APPLY).
• Sensory anchors: tissue rustle, fabric slide, hardware clinks (chains / clasps / buckles), leather/fabric grain under fingertips, foil glint.
• Setting: white unmade bed in natural window light; bathroom mirror in background for the outro held-up reveal; streetwear drops may use desk/floor.
• Close-up beat device: NOT a screen — product close-up. @Image2 is a product photo (or brand-site mobile view); the single finger-point lands on a hardware detail (chain, clasp, embossed logo).
• Dialogue character: hook is mystery tease — frame the unboxing as something the viewer doesn't yet know the contents of; do NOT name the product in the hook line. Arc: hook the unboxing mystery → brand name + drop context → reveal the material/silhouette while close-up holds on hardware → tactile/wearability insight (how it feels on the body) → punchline that invites the viewer to imagine themselves with the artifact.

APP_REVEAL

• When to use & why: SaaS, AI tools, mobile/web apps, agent-style products, dev tools, productivity tools — anything where the screen IS the product. Viewers convert when they see live UI doing the thing quickly; the close-up beat is the demo, the bookends are the social proof. Not pure hardware (→ TECH_UNBOX).
• Sensory anchors: micro-thumb gesture, brand-color highlight, UI alive with small motion, ambient room tone.
• Setting: cozy bedroom or couch POV; jeans/joggers at frame edges; warm window light.
• Close-up beat device: laptop on bed (desktop screenshot) or phone in hand (mobile screenshot — set mobile=true in step 3).
• Dialogue character: hook is bewildered curiosity — the creator can't categorize the thing yet, that's the point. Do NOT use feature lists or marketing language in the hook; lean into "I don't know what to call this" / "this is wild" register that makes the viewer wait for the name. Arc: bewildered hook → name the product + interaction model in human terms ("you just talk to it", "it builds X from Y") → reveal what it produces (concrete comma-separated examples) while close-up shows the page → personal-insight twist (what it replaces / changes in the user's workflow) → punchline + implicit/explicit "go try it" CTA.

FOOD_ASMR

• When to use & why: food brands, drinks, kitchen tools, snacks, restaurants with a takeout product — anything where the sensory peak (pour / sizzle / steam / first bite) carries the value prop. Viewers convert on hunger response — show the sensory peak, don't describe it.
• Sensory anchors: packaging rustle, knife-on-board, sizzle, pour stream, steam rising, satisfied exhale on the first bite.
• Setting: marble counter or warm wood kitchen, top-down framing.
• Close-up beat device: a product/dish close-up rather than a screen; phone in hand on the counter only if the brand has a delivery/recipe app.
• Dialogue character: hook is show-don't-tell — frame as a demonstration the viewer is watching unfold, not a description. The hook line lands while a hand or first ingredient is already in motion; the visual carries the curiosity. Arc: demonstration hook → name the product + first impression → narrate the sensory peak as it happens (pour / sizzle / steam) → satisfaction insight ("this is the new default") → punchline that hands off the recipe or shop link.

BEAUTY_APPLY

• When to use & why: skincare, makeup, cosmetics, fragrance, hair products, body care — anything where before/after + application ritual is the value prop. Viewers convert on visual transformation under matched lighting; symmetry between hook and outro is what sells the result as real. Not packaging-heavy luxury (→ HAUL_UNBOX).
• Sensory anchors: pump press, squeeze, glide on skin, glow lift, droplet beading, brush sweep.
• Setting: bathroom mirror, natural daylight or vanity lighting; same angle for the hook and the outro.
• Close-up beat device: a product close-up (bottle / tube / compact held in hand), not a screen.
• Dialogue character: hook is routine social proof — signal real use through a specific moment in the routine, not an invented duration. The hook plants the symmetry payoff that arrives in the after-shot. Arc: routine-use hook → name + source-backed key ingredient or claim → narrate the application as it happens (close-up of fingers/brush on skin) → after-shot reveal (same angle as hook) → punchline that signals exclusivity or repurchase intent.

FITNESS_TRANSFORM

• When to use & why: workout equipment, supplements, recovery tools, activewear, fitness apps with tracking — anything where the work-to-result transformation is the value prop. Viewers convert on relatable struggle followed by earned payoff — showing the protein-shake bottle is not enough, you have to show the workout.
• Sensory anchors: heavy breathing, scoop hitting powder, equipment click, sweat catching light, post-workout exhale.
• Setting: gym or home-gym; workout gear at frame edges; floor or bench level.
• Close-up beat device: phone in hand showing app stats / heart rate / time elapsed, OR product packaging close-up (scoop in jar, bottle pour).
• Dialogue character: hook is relatable resistance — name the struggle / friction / not-wanting-to ("I did NOT want to do this", "almost skipped today", "this was supposed to be a rest day"); earns trust by sharing the tired feeling before showing the work. Arc: resistance hook → name the product + non-numeric routine or source-backed claim → narrate mid-work moment while close-up shows the device or scoop → satisfaction insight that earns trust → punchline that frames continued use.

TECH_UNBOX

• When to use & why: gadgets, hardware, electronics, smart-home devices, wearables, peripherals, AI hardware (Framework laptop, AirPods, Whoop, Rabbit r1, Friend pendant, mechanical keyboards, ergonomic gear) — anything where the device + first-use moment is the value prop. The box ceremony signals premium positioning; viewers convert on seeing "does it actually work / what does it do" — the first-use beat is the conversion moment. Not HAUL (→ HAUL_UNBOX), not pure software/SaaS (→ APP_REVEAL).
• Sensory anchors: utility-knife slice, plastic peel, foam slide-out, power-on chime, tactile button press, haptic click, fan spin-up.
• Setting: wood desk, top-down framing during unbox; handheld during first-use; desk/lap context for ongoing use.
• Close-up beat device: the device itself once unboxed and powered on. @Image2 is typically a real photo of the device's screen at its key UI moment (first measurement, paired status, hero feature open); if the device has no screen, a clean hero photo of it mid-use.
• Dialogue character: hook is arrival ceremony — name that this is happening now ("just got this", "opening it"). Anticipation > description; the hook plants the question "what does it do?" that the first-use beat answers. Do NOT lead with specs. Arc: arrival hook → name + source-backed spec headline when claims_allowlist has one, otherwise qualitative headline → first-use reveal while close-up is on the device doing its thing → workflow-change insight ("this replaces / changes / fixes my X") → punchline that hands off where to find it without inventing cost or urgency.

6. Voice — model default

This skill has no voice-cloning input; the video model produces the spoken dialogue with its own default voice. No voice sample is fetched or passed. Proceed to step 7.

7. Generate — first attempt with the avatar, cartoonize on rejection, retry

After the Avatar-type probe in step 2.5 passes, attempt the call first with the avatar resolved in step 2 (caller-supplied or built-in fallback) exactly as-is. Only when seedance rejects the call do we restyle.

Retry budget: Seedance generation gets at most 2 total video attempts: the original avatar render in 7a, plus one cartoonized-avatar retry in 7c. Do not start a third Seedance render. If the second attempt fails, surface the 7d message or switch to Kling only when the user explicitly requested provider=kling or chooses it after the stop message.

7a. First attempt — avatar as-is

Call generate_reference_video:

• provider: seedance (default) or kling if user passed provider=kling
• aspect_ratio: 9:16 (default); 3:4 allowed only on seedance
• resolution: 720p (seedance only)
• duration: 15
• reference_images: [avatar_url, screenshot_url] (drop screenshot_url if step 3 failed)
• prompt: the multi-beat string from step 4
• sound: true (default — ambient + lip-sync produced by the model)

For provider=kling: convert the multi-beat prose into shots: [{prompt, duration}, ...] (5 shots × 3s = 15s sum), plus a top-level prompt summarizing the ad. References use <<<image_1>>> / <<<image_2>>> instead of @Image1 / @Image2.

If the call returns { task_id, status: "queued" }, poll task_status(task_id) using the async polling budget above until terminal (completed | failed | cancelled). On completed, capture result.url → video_url and proceed to step 8.

7b. On rejection — auto-cartoonize the avatar

If 7a returns 422 content_policy_violation on image_urls / reference_images (seedance + fal-queue moderation flags portraits that read as too photorealistic — even some Pixar-style 3D avatars get flagged), restyle the avatar in-place:

Call generate_image_edit:

• provider: "seedream" (native Pixar/3D-animated look)
• images: [avatar_url]
• aspect_ratio: same as the ad's aspect ratio
• resolution: "1K"
• watermark: false (seedream-only knob — keep the restyled avatar clean of provider watermark for the downstream lip-sync re-render)
• prompt: "Stylized 3D game character render — Unreal Engine 5 / Overwatch / Valorant / Apex Legends visual style. Anatomically grounded facial proportions with subtle stylization: slightly larger expressive eyes, defined sculpted cheekbone planes, smooth skin shader (smoother than photoreal, no micropore detail), idealized but believable features. PBR materials with subtle subsurface scattering, strand-based hair simulation, crisp cloth shader. Cinematic three-point studio lighting with strong rim light. Clearly a stylized AAA-game-character render — NOT photorealistic person, NOT Pixar plastic-toy cartoon, NOT exaggerated big-head proportions. Same person, same glasses, same outfit, same accessories. Centered medium portrait, neutral indoor background."

Capture returned URL → avatar_url_cartoon.

7c. Retry seedance with the cartoonized avatar

Re-run the exact same generate_reference_video call from 7a, swapping the avatar reference: reference_images: [avatar_url_cartoon, screenshot_url] (or [avatar_url_cartoon] if step 3 failed). All other params unchanged. Capture result.url → video_url.

7d. Final fallback — still rejected

If 7c also returns content_policy_violation, stop. Tell the user: the avatar reads as too realistic for seedance moderation even after auto-restyling; ask them to either supply a more stylized portrait themselves or rerun with provider=kling (kling has a separate moderation pipeline that accepts realistic avatars).

8. Deterministic brand/spec overlays

Before captions, add readable brand/spec copy with deterministic post-generation overlays. This is deliberately separate from the Seedance prompt because video-model text rendering corrupts brand wordmarks and spec labels.

Build overlay copy from Step 1:

{
  "brand_overlay_text": "<brand_name>",
  "grounded_spec_overlay_text": "<one short source-backed claim from claims_allowlist, or empty>",
  "overlay_font_color": "#111111 | #ffffff"
}

Rules:

• brand_overlay_text is always the exact brand_name from WebFetch. If brand_name is empty, use product_name; do not let the model invent a logo/wordmark.
• grounded_spec_overlay_text must be empty or one exact claim from claims_allowlist. Never overlay a number that is absent from claims_allowlist.
• Choose overlay_font_color for contrast against the generated frame, not brand aesthetics. If the screenshot/product surface is light or unknown, use #111111; use #ffffff only on clearly dark footage. Post-flight OCR must be able to read the overlay.
• Do not use this step for captions or per-word subtitles; Step 9 handles captions with one add_captions call.

Call edit_text_overlay once for the brand name:

edit_text_overlay(
  video_url: video_url,
  text: brand_overlay_text,
  position: "top_left",
  font_size: 54,
  font_color: overlay_font_color,
  start_s: 0.2,
  end_s: 15
)

Save the returned URL as brand_guarded_url. If grounded_spec_overlay_text is non-empty, call edit_text_overlay once more on brand_guarded_url:

edit_text_overlay(
  video_url: brand_guarded_url,
  text: grounded_spec_overlay_text,
  position: "top_right",
  font_size: 42,
  font_color: overlay_font_color,
  start_s: 6,
  end_s: 10
)

Save that returned URL as guarded_video_url. If no spec overlay is needed, set guarded_video_url = brand_guarded_url. If either overlay call returns { task_id }, poll task_status using the long-running task contract. If the deterministic overlay fails, stop and surface the failure instead of delivering a video whose only readable brand/spec text depends on Seedance.

Run frame-level OCR on the overlay window before captions. Video-level analysis can miss short-lived top-corner overlays, so extract one frame while both overlays should be visible:

extract_frame(
  video_url: guarded_video_url,
  time_s: grounded_spec_overlay_text ? 7 : 1
)
# Save returned url as overlay_qa_frame_url

analyze_media(
  media: overlay_qa_frame_url,
  query: "Expected brand text: ${brand_overlay_text}
Expected spec text: ${grounded_spec_overlay_text}
Return JSON only: { \"visible_text_ocr\": string[], \"brand_visible\": boolean, \"spec_visible\": boolean, \"observations\": string[] }. OCR/read all visible text. brand_visible is true only if the exact expected brand text above is readable. spec_visible is true if expected spec text is empty or the exact expected spec text above is readable."
)

If brand_visible is false, or spec_visible is false when grounded_spec_overlay_text is non-empty, stop and rerender the deterministic overlay once with the opposite overlay_font_color (#111111 ↔ #ffffff). Re-extract overlay_qa_frame_url and rerun the same OCR check. If the retry still fails, stop and surface overlay_qa_frame_url and the OCR JSON; do not proceed to captions.

9. Captions — single-shot styled burn (default on)

If variants is requested, do not run the single-output add_captions call in this step. Leave guarded_video_url uncaptioned and continue to Aspect-ratio variants, where each requested aspect is reframed first and captioned once after reframe.

If variants is absent and captions=false, skip caption generation and set final_url = guarded_video_url. Otherwise, use one add_captions call instead of chaining edit_text_overlay per chunk — much faster (≤5 min single call vs 5–8 min sequential), and the styles position captions correctly out of the box.

Call add_captions:

• video_url: guarded_video_url from step 8
• style: "tiktok" (default — word-by-word purple highlight, Bebas Neue, all caps, rendered at the bottom of the frame; classic TikTok-creator look that keeps the face and screen clear). Alternatives: "hormozi" (lower-middle yellow highlight, more aggressive — overlays part of the phone-in-hand close-up beat), "classic" (plain bottom subtitle bar, safest), "karaoke" (progressive color fill, also bottom).
• font_size: 60 — overrides the per-style default; tuned for 9:16 readability without dominating the frame.
• language: pass the BCP-47 code for the page language detected in step 1 ("en", "zh", "ja", "es", etc.) — skips auto-detect and avoids misrouting CJK to a Latin-only font path.

Capture the returned URL → final_url.

Aspect-ratio variants

Use optional variants=9:16,16:9,1:1 when the user wants vertical short-form, landscape, and square feed outputs from the same UGC ad. variants overrides aspect_ratio to native 9:16; aspect_ratio=3:4 is single-output only and must not be combined with variants. If the user requests variants with aspect_ratio=3:4, use the 9:16 native variant flow and state that 3:4 applies only to single-output runs.

Call generate_reference_video once for the native 9:16 ad, then run Step 8 deterministic overlays once to produce guarded_video_url. Do not re-run product fetch, avatar analysis, screenshot capture, avatar cartoonization, Seedance/Kling generation, deterministic overlays, or any other expensive provider call for extra variants.

Do not run the single-output add_captions call from step 9. Build raw variant_sources first, then build a flat variant_urls object:

{
  "9:16": "<native final_url>",
  "16:9": "<edit_reframe url>",
  "1:1": "<edit_reframe url>"
}

• For 9:16, set variant_sources["9:16"] = guarded_video_url.
• For 16:9 and 1:1, call edit_reframe(video_url=guarded_video_url, target_aspect="<aspect>", fill_mode="blur") and save each returned URL into variant_sources. Blur fill preserves the full native vertical ad over a blurred background instead of cutting off the creator or captions.
• If captions=true, caption after reframe: run add_captions on each variant_sources value so burned captions are positioned for that aspect, then save those captioned URLs into variant_urls. If captions=false, save the uncaptioned native and reframed URLs directly into variant_urls.
• Set final_url = variant_urls["9:16"] so the primary deliverable matches the native requested variant.
• Treat edit_reframe and per-variant caption burn as the cheap final composite / reframe stage. If a reframe fails, return the successful variant URLs plus the failed aspect and tool error; do not resubmit the expensive generation.

Heads up — variants use blur fill. 16:9 and 1:1 keep the full native 9:16 ad visible over a blurred background. Keep the creator, product, and captions centered for readability, but do not describe these outputs as cropped variants.

10. Return

Return final_url on one line, plus a one-line summary: which category ran, whether the avatar was caller-supplied / built-in fallback / cartoonize-recovered, whether the screenshot was used or fell back to prose, the provider chosen, the language detected for dialogue, and whether captions were burned on. If variants was requested, include the flat variant_urls object in the same response.

Post-flight quality gate

Before declaring success, call analyze_media on final_url when captions were burned, or on guarded_video_url when captions=false, and ask for a structured verdict. If variants was requested, run the same gate on each variant_urls value and key any warning by aspect ratio.

Before calling analyze_media, substitute concrete expectations into the query: include exact brand_name, exact product_name, and the actual claims_allowlist values as a JSON array. Do not leave placeholders for the model to infer.

Expected brand: ${brand_name}
Expected product: ${product_name}
Allowed numeric/spec claims: ${JSON.stringify(claims_allowlist)}

Return JSON only: {
  "verdict": "clean" | "degraded" | "catastrophic",
  "visible_text_ocr": string[],
  "unauthorized_numeric_claims": string[],
  "observations": string[],
  "quality_warning": string | null,
  "re_roll_suggestion": string | null
}
OCR / read all visible text. Use the concrete expected values above when checking that `brand_name` and `product_name` are spelled correctly anywhere they appear, deterministic brand/spec overlays are legible, captions are present when requested and not garbled, and no visible numeric claim appears unless it exactly matches one of the actual `claims_allowlist` values above. Also check that the avatar / creator remains visually stable, the product screenshot or prose fallback matches the product page, and there are no black frames or wrong-product shots.

• If verdict is clean, return final_url normally.
• If verdict is degraded, return final_url plus the quality_warning so the user can review before publishing.
• If verdict is catastrophic, if visible_text_ocr contains garbled brand/spec text, if brand_name is misspelled, or if unauthorized_numeric_claims contains any claim not in claims_allowlist, do not call the ad complete; surface the verdict and re_roll_suggestion instead of declaring success.

Failure modes

Load-bearing phrases

These anchors keep the ad from drifting into a generic product demo:

Examples

• /pika:ugc-ads https://pika.me avatar_url=https://cdn/face.png → APP_REVEAL, 9:16, seedance, real screenshot, captions on
• /pika:ugc-ads https://maisonbrune.com avatar_url=https://cdn/face.png aspect_ratio=3:4 → HAUL_UNBOX, 3:4, seedance
• /pika:ugc-ads https://pika.me avatar_url=https://cdn/face.png provider=kling captions=false → APP_REVEAL, 9:16, kling shots[], no captions
• /pika:ugc-ads https://pika.me → no avatar_url → uses the built-in fallback Pixar-style female creator portrait, runs end-to-end