main-image-pipeline

name

main-image-pipeline

description

End-to-end main image optimization pipeline — the flagship CRO play. Pulls SellerApp competitor SERP, generates 8 main image concepts via Higgsfield (GPT Image / Nano Banana Pro), runs a 3-second billboard test on top 3 via ProductPinion, then runs a final Image Split Test to declare a winner. Use when running `main-image-pipeline {ASIN}` or when an ASIN has a CTR problem (low CTR vs category index). Replaces the 3-week external designer cycle with a same-day-to-3-days pipeline.

main-image-pipeline — Flagship CRO Pipeline

Single command that orchestrates the highest-ROI play in the skill library: research → AI concepts → billboard test → split test → ship-ready winner. Main image is the highest CTR lever on Amazon; this is the one to run first when an ASIN has a CTR gap.

Methodology — read before rendering

Before generating ANY main-image concept, read reference/02-visual-content/main-image-creative-director.md in full. That file is the source of truth for:

• The 8 enhancement techniques (one per concept, no repeats across the variants)
• The 6-section mandatory prompt structure (Product Position / Angle / Lighting / Extra Items / White Background / Square Format)
• The 5 thumbnail rules (frame fill 85%, hero angle, perceived quality, pattern interrupt, instant category recognition)
• The 200–350 word prompt cap + the terminator anchor
• The 4-axis scoring rubric (40/30/15/15 — F/I/C/R) and 5 binary pre-screen checks
• Reference image handoff (live product photo from brain/products/{asin}-{geo}.json, always passed)
• Aspect ratio enforcement: 1:1 hard-coded for main images
• Iteration loop: cap at 3 attempts per concept, fail honestly if the model can't deliver

Aspect ratio is non-negotiable. Every prompt ends with square 1:1 format, e-commerce product photography, hyper-realistic, ultra-sharp, 8k. Wrong-ratio outputs trigger automatic re-generation.

Concept gallery output emits florence-concepts-{asin} per the artifact protocol — see 0-paste-this-into-custom-instructions.txt § When to emit per-product artifacts.

Prerequisites

This pipeline calls 3 skills in sequence, so all of these must be available:

• ✅ SellerApp via n8n MCP (workflow 9RmjDT107uXtrImf)
• ⚠️ Higgsfield MCP — required for Phase 2; see prerequisites in main-image-concepts
• ⚠️ ProductPinion MCP — required for Phases 3-4; see prerequisites in main-image-poll

If any are missing, skill stops at that phase with a setup instruction and a --resume-from={phase} flag for after the user connects the missing piece.

Invocation

main-image-pipeline {ASIN}
main-image-pipeline {ASIN} --skip-3-second    # skip Phase 3, go straight to split test
main-image-pipeline {ASIN} --concepts=10      # default 8
main-image-pipeline {ASIN} --resume-from=3    # if a prior run stopped at a phase

Output

Master report: /tmp/cro-content/{ASIN}-main-pipeline-{date}.md — links to each phase output. Final ship-ready image saved to /tmp/cro-content/{ASIN}-main-WINNER.png.

State File

Persists state to /tmp/cro-content/{ASIN}-main-pipeline.state.json after every phase so the pipeline is resumable across sessions.

Phase 0 — Re-entry Check

Check: /tmp/cro-content/{ASIN}-main-pipeline.state.json

If found:

• "Found existing pipeline at Phase {N}. Resume or restart?"
• Resume → skip to phase
• Restart → archive old state, begin Phase 1

Phase 1 — Research & SERP Sweep

Calls competitor-sweep {ASIN} (which itself calls asin-deep-research if no brief exists).

Outputs needed for Phase 2:

• Top 3 purchase drivers (visual must-shows)
• Top 3 objections (preempt list)
• Customer demographic
• Competitor SERP cluster — the dominant visual style your thumbnail must break from
• Quality benchmarks (target rating, image count, etc.)

State after Phase 1: phase: "concepts", brief_path: "/tmp/cro-research/..."

Phase 2 — AI Concept Generation

Calls main-image-concepts {ASIN} --research={brief_path} --count=8.

Default models: nano-banana-pro primary, gpt-image fallback. Best-only — no multi-model spread.

Outputs:

• 8 generated concepts in /tmp/cro-content/{ASIN}-main-concepts/
• Scored concept grid (6-dim rubric)
• Top 3 recommended

State after Phase 2: phase: "billboard_test", top_3: ["A", "B", "C"]

Phase 3 — 3-Second Billboard Test (ProductPinion)

The "Show Me, Don't Tell Me" rule from MASTER-CRO-REFERENCE.md requires the main image to communicate in <2 seconds. This phase runs ProductPinion's Pinion Ask 3 Second Test on the top 3 concepts to validate before the more expensive split test.

Setup:

• Test type: Pinion Ask → 3 Second Test (per ProductPinion docs)
• Each concept shown for 3 seconds, followed by quick recall questions:
  • "What product did you see?" (free text)
  • "What stood out?" (free text)
  • "Would you click to learn more?" (yes/no)
• Sample: 50 shoppers per concept (3 concepts = 150 total)
• Audience: research-brief-derived demographic

Pass criteria per concept:

• ≥80% correctly identified the product category
• ≥40% would click

Concepts that fail this gate are dropped from Phase 4. If 0-1 pass: stop pipeline, return to Phase 2 with feedback.

State after Phase 3: phase: "split_test", finalists: ["A", "C"] (those that passed the gate)

Phase 4 — Image Split Test (ProductPinion)

Calls main-image-poll {ASIN} --concepts={finalists} with sample size of 100 per option (200+ total).

Outputs:

• Winner declaration with % preference + confidence
• Qualitative "why" themes
• Ship/iterate recommendation

State after Phase 4: phase: "complete", winner: "A", winner_path: "/tmp/.../concept-A.png"

Phase 5 — Ship-Ready Output

Master report at /tmp/cro-content/{ASIN}-main-pipeline-{date}.md:

# Main Image Pipeline — {Title}

**ASIN:** {ASIN} | **Date:** {date} | **Total run time:** {duration}

## TL;DR

🏆 **Winner:** Concept {X} — {brief description}
**Confidence:** {high/med/low} from {N} shoppers
**Recommended next step:** Launch in Amazon Manage Your Experiments

## Pipeline Trace

| Phase | Output | Link |
|-------|--------|------|
| 1. Research | Brief + competitor matrix | [link] |
| 2. AI Concepts | 8 generated, top 3 picked | [link] |
| 3. Billboard Test | 2 of 3 passed gate | [link] |
| 4. Split Test | Winner declared | [link] |

## Winner Image

![](concept-A.png)

## Why It Won (qualitative)

- Theme 1: {%}
- Theme 2: {%}
- Theme 3: {%}

## Designer Brief (handoff to polish)

{For external designer if final cleanup is needed:}
- Use the AI-generated concept as composition reference
- Maintain: {key elements that drove the win}
- Polish: {color accuracy / cutout cleanup / lighting tweak}
- Final spec: 2000×2000 PNG, white #FFFFFF background, sRGB

## MYE Launch Plan

- Test: New main image vs current
- Duration: 14 days minimum (per `05-testing/test-design-methodology.md`)
- Success metric: CTR lift ≥10% with 95% confidence
- Failure plan: revert to current, queue Concept {Y} for next test

Reference Files

• This skill orchestrates asin-deep-research, competitor-sweep, main-image-concepts, main-image-poll
• ~/.claude/skills/cro/main-image-best-practices.md
• Vault: CRO-Knowledge-Base/02-visual-content/main-image.md
• Vault: CRO-Knowledge-Base/05-testing/test-design-methodology.md

Quality Bar

• All 4 phases completed (or graceful stop at MCP-prerequisite phase)
• State file persisted after each phase (resumable)
• Master report has winner, confidence, "why", and MYE launch plan
• Designer brief section is filled (this is the handoff to external polish)
• Winner image saved to {ASIN}-main-WINNER.png for easy retrieval

Failure Modes

Phase	Failure	Action
1	Research returns thin data	Stop, prompt user; this is the foundation
2	All 8 concepts fail compliance	Re-run with stricter prompt template
3	0-1 concepts pass billboard	Return to Phase 2 — concepts not strong enough
4	No statistical winner (e.g. 51/49)	Two paths: ship best, OR run a 3rd contender

Auto-Triggers

Runs (without prefix) when:

• User says "fix the main image for {ASIN}" / "low CTR on {ASIN}"
• Diagnostic shows low CTR + healthy CVR (per 04-data-analysis/metric-to-action-framework.md)
• User asks for "the full main image play" or "end-to-end main image"

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.