main-image-concepts

name: main-image-concepts description: Generate 5-10 Amazon main image concepts for an ASIN using Higgsfield AI image generation. Defaults to GPT Image (latest) and Nano Banana Pro — best-quality only. Each concept is scored against the 6-dimension main image rubric (Fidelity, Background, Scroll-Stop, Compliance, Creative, Quality). Use when running main-image-concepts {ASIN}, when you need standalone main image ideation without polling, or as a sub-step of main-image-pipeline.

main-image-concepts — AI-Generated Main Image Concepts

Standalone concept generation using Higgsfield. Replaces the early "designer ideation" cycle (5-10 days, $500-$2000) with same-day AI concepts that go to designers for polish.

Methodology — read before rendering

Read reference/02-visual-content/main-image-creative-director.md in full before generating any concept. That file is the source-of-truth for: the 8 enhancement techniques (one per variant, no repeats), the 6-section mandatory prompt structure, the 5 thumbnail rules, the 4-axis scoring rubric (40/30/15/15), and the iteration loop.

Hard rules from that file (non-negotiable):

• Aspect ratio: 1:1, locked in at start AND end of every prompt
• Reference image always: pull live product photo from brain/products/{asin}-{geo}.json and pass to Higgsfield's image-to-image input
• Brand guidelines respected: read brain.business.brand_guidelines for colours / fonts / visual style keywords; if empty, fall back to neutral premium aesthetic and flag on the concept card
• Cap at 3 attempts per concept before surfacing failure honestly

Output emits florence-concepts-{asin} per the artifact protocol.

Prerequisites

Higgsfield MCP must be connected at https://mcp.higgsfield.ai/mcp. If not connected:

1. Tell user: "Higgsfield MCP isn't connected. Open Claude Settings → Connectors → Add Custom Connector → URL https://mcp.higgsfield.ai/mcp → sign in with Higgsfield account."
2. Pause until connected, then resume.

Invocation

main-image-concepts {ASIN}
main-image-concepts {ASIN} --count=8           # default 8
main-image-concepts {ASIN} --model=gpt-image   # or nano-banana-pro (default), or both
main-image-concepts {ASIN} --research={path}   # use existing research brief

Output

Writes to /tmp/cro-content/{ASIN}-main-concepts-{date}.md with embedded image URLs and a scoring grid. Saves images to /tmp/cro-content/{ASIN}-main-concepts/ directory.

Default Models (Best-Only Policy)

Per the skill plan: best-quality models only, no multi-model spread by default.

• Primary: nano-banana-pro — best for product fidelity + clean Amazon look
• Alternate: gpt-image (latest) — best for creative angles + dramatic lighting
• Fallback: flux-kontext for product replacement / packaging variations

To explicitly compare across models, use main-image-multi-model instead.

Phase 1 — Source the Brief

Required inputs (in order of preference):

1. --research={path} — existing brief from asin-deep-research
2. Existing /tmp/cro-research/{ASIN}-research-brief-*.md (most recent)
3. If neither: trigger asin-deep-research first, then resume

The brief provides:

• Top 3 purchase drivers (what to lead with visually)
• Top 3 objections (what to preempt)
• Customer demographic (who's in the image, if person)
• Competitor SERP cluster (what visual styles are saturated)

Required: confirm with user the product reference image URL (clean PNG, white BG) before generating. Higgsfield Soul/Nano Banana need this for fidelity.

Phase 2 — Concept Brief Generation

Translate the research into 8 distinct concept directions per 02-visual-content/main-image.md:

Standard concept slate (pick 5-8 from this menu based on research):

Each concept gets a structured prompt that includes:

• Product fidelity instruction (use reference image)
• Background rule: pure white #FFFFFF
• Composition: product fills 85%+
• Compliance: no text/logos/badges (Amazon main image rule)
• Style: photorealistic, e-commerce, soft natural lighting
• Aspect ratio: 1:1 (Amazon main image)
• Resolution: 2000×2000 minimum

Phase 3 — Generate via Higgsfield MCP

For each concept (8 by default):

• Call Higgsfield generate with primary model (nano-banana-pro)
• Pass: prompt, reference image, aspect 1:1, resolution 2000×2000

Run in batches of 4 (Higgsfield can handle parallel; respect MCP rate limits).

If a generation fails the fidelity check (product visibly different), re-run with gpt-image as fallback.

Phase 4 — Score Against the 6-Dimension Rubric

Per MASTER-CRO-REFERENCE.md §3 (Main Image scoring):

Pass: average ≥7 with no dimension <5. Auto-fail triggers: product too small, flat catalog shot, dull lighting, cluttered, visible cutout artifacts, product blends into white.

Phase 5 — Output

# Main Image Concepts — {Title}

**ASIN:** {ASIN} | **Date:** {date} | **Models:** {models} | **Reference:** {ref-image-url}

## Brief Summary
- Top driver: {from research}
- Top objection to preempt: {from research}
- SERP visual cluster: {from competitor sweep}
- Differentiation play: {strategy chosen}

## Concept Grid

| # | Concept | Image | Fid | BG | Stop | Compl | Creat | Qual | Avg | Pass |
|---|---------|-------|-----|-----|------|-------|-------|------|-----|------|
| 1 | Hero, dramatic angle | ![](path) | 9 | 10 | 8 | 10 | 8 | 9 | 9.0 | ✅ |
| 2 | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... |

## Top 3 Recommended (for next step)

1. **Concept #X — {name}** (Avg: 9.2) — Why: {1 line on which research signal it lands}
2. **Concept #Y — {name}** (Avg: 8.8) — Why: {...}
3. **Concept #Z — {name}** (Avg: 8.5) — Why: {...}

## Recommended Next Skill

- Validate with shoppers: `main-image-poll {ASIN} --concepts={1,2,3}`
- Or run the full pipeline: `main-image-pipeline {ASIN}` (research → generate → 3-second test → poll → winner)

Reference Files

• ~/.claude/skills/cro/main-image-best-practices.md — full 6-dim scoring + Amazon compliance
• ~/.claude/skills/hero-image/prompt-engineering.md — proven Higgsfield prompt patterns
• ~/.claude/skills/hero-image/image-scoring.md — scoring rubric details
• Vault: CRO-Knowledge-Base/02-visual-content/main-image.md
• Vault: CRO-Knowledge-Base/MASTER-CRO-REFERENCE.md §3 (Main Image)

Quality Bar

• At least 5 distinct concept directions (not 8 variations of one idea)
• Every concept passes Amazon compliance (no text/logos)
• Reference image used — products are recognizable
• Scoring filled for every concept (no blank cells)
• Top 3 picks justified with research-signal references
• Saved images accessible at /tmp/cro-content/{ASIN}-main-concepts/

Auto-Triggers

Runs (without prefix) when:

• User asks "generate main image concepts for {ASIN}"
• User completes asin-deep-research and asks for visuals
• main-image-pipeline calls it as a sub-step

v0.1.12 — Read brain.image_strategy before generating

Before drafting any Higgsfield prompt, read brain.image_strategy from working memory:

• If null → pause and offer image-strategy first (~10 min, makes every future render bespoke). If user proceeds without, flag concept cards with "No category research used — generic aesthetic."
• If set + fresh (<90d) → inject prompt_adjustments.scene_keywords into prompt section 1, palette_keywords + mood_keywords into section 4, and do NOT include {anti_patterns_csv} near the end. Cite the strategy on each concept card.
• If stale (>90d) → flag and recommend image-strategy --refresh.

This brand-level strategy comes from skills/image-strategy.md's top-15-bestsellers analysis. Same adjustments apply across every render for this brand.

v0.1.13 — Visual verification gate + base64 embedding (NON-NEGOTIABLE)

Florence does NOT present an image she hasn't actually looked at, AND she does NOT use raw Higgsfield URLs in artifact HTML. Two blocking rules added in v0.1.13:

Verification gate

After Higgsfield returns each generated image, BEFORE adding to any artifact OR sending to Pinion:

1. Load the image into Florence's multimodal context — paste the URL in chat so Cowork's multimodal Claude loads it natively. Narrate: "Looking at #{N} before I include it."
2. Visually inspect against: aspect ratio (1:1 main+listing / 16:9 A+), product fidelity vs reference, technique landed (visibly), background appropriate, no clipart leak (anti-clipart rules), no text-on-main-image (TOS), no model faces.
3. If anything fails → re-prompt with specific fix + regenerate. Cap at 3 attempts per concept; surface honestly on attempt 4.
4. Only after ALL concepts pass → proceed to artifact emission.

The eye trumps the score. Even if the rubric said 100/100, if visual inspection finds a wrong product or off-aspect output, it fails the gate.

Base64 embedding

Higgsfield URLs are temporary AND Cowork's artifact iframe sandbox blocks external image loads in many builds. For every verified concept:

1. HTTP GET the Higgsfield URL → fetch image bytes
2. Detect MIME type from response headers (typically image/png)
3. Base64-encode the bytes
4. Substitute {{image-src}} (or {{winner-image-src}} for tests) with data:image/png;base64,<encoded> — NOT the raw Higgsfield URL

This makes the artifact self-contained — survives sandbox + URL expiry. ~1-3 MB per image is fine for Cowork.

The live product image ({{product-image-url}} in concepts.html hero strip / dossier head / cockpit product cards) stays as the live Amazon CDN URL — that's permanent and not affected.