Product URL
https://example.com/products/skin-recovery-serum
Used for offer, claims, ingredients/features, social proof, and landing-page CTA extraction.
Ivory Signal / Prompt-Reference Lab MVP
Plan the ad before spending render credits.
This no-video workflow turns product inputs, competitor/source placeholders, ordered references, hooks, storyboard beats, and prompt constraints into a demoable creative package.
Generation is intentionally disabled. The lab shows what Bloom Video can structure and verify before a provider is connected or a credit is spent.
Provider spend disabled Ordered @image references Evidence placeholders
Lab status
5
scan inputs
4
reference slots
0
provider credits
Render switch: off
The MVP outputs planning artifacts only: source pack, hooks, storyboard, reference order, and prompt recipe.
Step 1 / Product scan
Inputs become constraints, not loose notes.
Product image
Hero bottle on matte bathroom counter
Anchors packaging, color, form factor, and product visibility requirements before any render.
Audience
Busy women 28–44 with sensitive skin and decision fatigue
Keeps hooks tied to a specific buyer pain instead of generic beauty copy.
Brand tone
Clinical but warm; proof-first; no miracle claims
Feeds policy/safety notes and prompt recipe constraints.
Target platform
TikTok/Reels 9:16, 15 seconds
Sets pacing, text density, shot count, and planned render-cost estimate.
Step 2 / Source pack
Evidence scaffolding before generation.
The first MVP uses placeholders for competitor ads, review pain, and trend links so the workflow is demoable without scraping or paid provider calls.
Competitor ad library
Meta/TikTok ad examples placeholder
Pattern: creator-style sink demo, close-up texture shot, one hard before/after claim to avoid.
Review-pain excerpts
Product/retailer reviews placeholder
Repeated pain: sticky finish, confusing routines, irritation after trying too many actives.
Trend/source links
Saved social trend URLs placeholder
Format: 3-shot problem/ritual/payoff arc with captions that can be read muted.
Cost/pricing baseline
Seedance-style 720p/15s planning estimate
No render is launched. Lab shows the planned scene count and spend range before credits are used.
Step 3 / Ordered reference_set
Every reference gets a stable slot.
Prompt recipes must preserve image order so @image1, @image2, @image3, and @image4 stay auditable across storyboard beats and QA checks.
#1 @image1
Product packshot
Packaging truth source
Bottle shape, label zone, and cap color must stay recognizable.
#2 @image2
Texture macro
Ingredient/finish reference
Cream/serum texture should look lightweight, not sticky or oily.
#3 @image3
Creator bathroom setup
Scene and lighting reference
Use warm daylight, handheld creator framing, and clean counter composition.
#4 @image4
Audience lifestyle frame
Human context reference
Show routine simplification without implying medical results.
Step 4 / Hooks and storyboard
Concepts with source refs and scene beats.
Cost-control hook
“Stop paying for AI ads before the idea is proven.”
Leads with the credit-opacity complaint and positions Bloom Video as the preflight lab.
Refs: credit opacity / failed-generation credit burn / cost preview
Buyer-pain hook
“Your skincare routine should not feel like a chemistry exam.”
Turns review-pain signals into a simple before/after narrative without medical claims.
Refs: review-pain excerpts / brand tone / policy guardrail
Competitor-gap hook
“A calm serum ad that shows the product for more than half a second.”
Responds to generic beauty ads where the product is lost behind creator montage pacing.
Refs: competitor ad library / product packshot / QA product visibility
0–3s
Cold-open sink counter with product visible beside a cluttered routine.
Use @image1 for packaging and @image3 for bathroom lighting. Add caption: “Too many steps?”
VO: If your routine got complicated before your skin got calmer…
3–8s
Macro texture swipe, then creator removes two redundant bottles from frame.
Use @image2 for texture. Keep text minimal and readable on mobile.
VO: …build the ad around the real buyer pain: decision fatigue, sticky formulas, and overpromising.
8–13s
Product in hand, clean counter, simple three-word benefit overlay.
Use @image4 for audience context. Avoid medical/guaranteed-result language.
VO: Show the product clearly, keep the promise grounded, and preflight cost before rendering.
13–15s
End card with product, offer, and QA checklist badge.
Static finish with caption-safe CTA and source-backed concept badge.
VO: Bloom Video turns references into ads with receipts.
Step 5 / Prompt recipe
Output is copyable, auditable, and no-video.
Mode: Seedance-style image-to-video planning only
Rendering enabled: no
Credit note: Planning: 0 provider credits. Render preview: 3 scenes x 5s, estimate pending provider connection.
Prompt
Create a 15s vertical product ad using @image1 as product packaging truth, @image2 for serum texture, @image3 for bathroom lighting/composition, and @image4 for audience lifestyle context. Keep product visible in every scene, use warm clinical tone, avoid medical-result claims, include readable muted captions, and follow the storyboard beats in order.Negative prompt / QA guardrails
No fake before/after cure claims, no unreadable text, no warped label, no extra fingers, no hidden product, no over-glossy plastic skin, no unsupported dermatologist claim.Step 6 / Gemini QA contract
Every generated video gets a scored report before delivery.
Default QA is Gemini 3.1 Flash-Lite. Calls are stubbed for this phase, with explicit labels for Gemini 3.5, OpenAI, and human billing escalation when retry/refund decisions need stronger evidence.
Prompt adherence
20%Pass: Storyboard order, reference roles, duration, platform, and tone match the prompt recipe.
Retry: Scene order drifts, target audience/tone is generic, or a required reference is weakly followed.
Refund: Output ignores the core prompt or returns an unrelated video.
Product visibility
20%Pass: Product remains recognizable in every planned scene and the packshot truth source is preserved.
Retry: Product disappears, label warps, or the hero shot is too brief for ad use.
Refund: Product identity is unusable or replaced with a different item.
Text legibility
15%Pass: Muted captions and end-card text are readable on mobile.
Retry: Captions are partially garbled, cropped, or too dense.
Refund: Critical CTA/offer text is unreadable across the clip.
Brand safety
20%Pass: Claims stay grounded, policy-safe, and compatible with negative-prompt guardrails.
Retry: Borderline wording, visual implication, or brand-tone mismatch needs rerender/edit.
Refund: Unsupported medical/legal/safety claim appears despite explicit guardrails.
Artifact quality
20%Pass: Motion, anatomy, object continuity, and scene composition are usable for a draft ad.
Retry: Visible artifacts hurt polish but the concept remains salvageable.
Refund: Corrupt file, severe warping, or unusable motion makes the render non-deliverable.
Cost & policy
5%Pass: QA is stubbed/no-spend by default and any paid escalation is labeled before use.
Retry: A rerender should be free because failure came from model/provider quality rather than user change.
Refund: Billing/manual review is required when paid output fails the contracted QA threshold.
Escalation ladder
Fast default, stronger review only when needed.
Default QA
Gemini 3.1 Flash-Lite
Every generated video gets the fast, cheap first-pass scoring report. Current implementation is stubbed and spends no model/provider credits.
Borderline rerender review
Gemini 3.5
Use for major retry-eligible issues, expensive rerender decisions, or disputed prompt-adherence scores.
Independent second opinion
OpenAI video QA
Use when customer feedback conflicts with Gemini QA or a high-value account needs independent evidence.
Billing/manual review
Human review
Use for refund recommendations, policy-risk outputs, and any billing decision that should not be automated.