image-creator（图像创作者）

Image Creator

Turn post copy into brand-aligned visuals. One tool (nanobanana), one output (image files). This is the final step in the creative-mkt pipeline: post-writer produces copy → image-creator produces matching visuals.

Purpose

Great copy without matching visuals gets scrolled past. Image Creator bridges that gap — it reads your post copy (or visual direction notes), loads your brand's visual identity, engineers a precise prompt, and generates images through nanobanana MCP.

The skill focuses on generation only. Prompt pattern knowledge lives in references/prompt-patterns.md — read it during Step 3 for detailed templates and examples.

Memory Auto-Load Protocol

Run this before any work:

1. Check creative-memory/ exists → if not, create from creative-memory-template/
2. Load creative-memory/visual-guidelines.md  ★ CRITICAL ★
   → Extract: Color Palette (HEX codes), Image Style, Typography feel, Do's & Don'ts
   → If empty or missing: warn user, suggest running visual-extractor first
   → Proceed anyway with generic professional style as fallback
3. Load brand-memory/voice-profile.md (read-only)
   → Map brand personality traits to visual mood (e.g., "bold & direct" → high contrast, strong lines)
4. Load brand-memory/positioning.md (read-only)
   → Note differentiation angle for visual storytelling
5. Load creative-memory/content-examples.md (read-only)
   → Check visual direction notes from past high-performing posts

Access rules:

Input Gathering

Collect from the user:

If the user provides a post-writer output file, extract:

• The post copy (for message analysis)
• The Visual Direction section (if present — post-writer attaches these)
• The platform (from file metadata or ask)

Step 1: Message + Tone Analysis

Extract visual direction from the copy. This step turns words into visual concepts.

1a. Core message extraction Summarize the post's central idea in one sentence. This becomes the image's subject.

1b. Emotional tone mapping Map the copy onto these spectrums:

The position on each spectrum informs color temperature, composition tightness, and visual complexity.

1c. Visual keyword extraction Derive 3–5 concrete visual concepts from the message + tone:

• Abstract concepts → tangible metaphors (e.g., "growth" → ascending stairs, sprouting plant)
• Emotional tone → visual treatment (e.g., "trust" → clean lines, blue tones, open space)
• Brand context → style constraints (from visual-guidelines)

1d. Brand alignment check Cross-reference visual keywords against visual-guidelines.md:

• Do the keywords align with the brand's Image Style direction?
• Are there conflicts with the Don'ts list?
• Adjust keywords to stay within brand boundaries while serving the copy's message.

Step 2: Platform Configuration

Select the right technical settings based on the target platform.

Platform defaults:

Model selection logic:

• Default: auto (nanobanana decides)
• Force pro when: brand integrity is critical, text-in-image needed, portfolio/hero usage
• Force flash when: high volume batch, stories/ephemeral content, quick iterations

Composition guidance per ratio:

• 1:1 — Centered subject, symmetric balance, ample breathing room
• 4:5 — Vertical emphasis, subject in upper 60%, space for caption overlap
• 9:16 — Strong vertical flow, stacked elements, text zones at top and bottom thirds
• 16:9 — Horizontal storytelling, rule-of-thirds subject placement, cinematic feel

Step 3: Prompt Engineering

This is the core value of the skill. Build the generation prompt using a structured formula.

Read references/prompt-patterns.md now — it contains detailed templates, style libraries, and worked examples for each use case.

Prompt formula:

[Subject/Scene] + [Style Directive] + [Color/Mood] + [Composition] + [Platform Specs] + [Negative Prompt]

Building each component:

Subject/Scene — Convert Step 1's visual keywords into a concrete scene description. Be specific: "a person" is weak; "a focused entrepreneur working at a clean desk with morning light" is strong. Use the specificity ladder in prompt-patterns.md.

Style Directive — Pull from visual-guidelines.md Image Style section:

• Photography-first brands → specify camera angle, lighting type, depth of field
• Illustration-first brands → specify illustration style, line weight, flatness
• If no guidelines exist → default to "clean, modern, professional photography style"

Color/Mood — Embed brand colors directly as HEX values in the prompt. This is more effective than color names alone:

• "color palette dominated by deep navy (#1A2B4C) with warm gold (#C9A96E) accents"
• Pull Primary + Accent from visual-guidelines

Composition — Use Step 2's platform-specific composition guidance

Platform Specs — Aspect ratio from Step 2

Text Overlay Language — If the user specified a language, all text overlays in the image (headlines, captions, CTAs) must be written in that language. If no language was specified, default to English.

Negative Prompt — Always include:

• Universal: blurry, low quality, watermark, distorted text, artifacts, oversaturated
• Brand Don'ts from visual-guidelines (convert to negative prompt form)
• Any user-specified exclusions

Variation strategy — For 2–3 variations, modify ONE element per variation:

• Variation A: base prompt (primary interpretation)
• Variation B: shift composition (different angle or framing)
• Variation C: shift color emphasis (accent color more prominent)

Step 4: Generate Images

Call nanobanana MCP to create the images.

nanobanana capabilities:

Generation sequence:

1. Construct the full prompt from Step 3 components
2. Call generate_image with prompt + aspect ratio + model selection
3. Repeat with modified prompts for each variation
4. Review each generated image immediately:

- Brand colors reflected? - Aspect ratio correct? - No unintended text or artifacts? - Mood matches the copy's tone?

1. If a result is off → adjust the prompt (usually the subject description or style directive) and regenerate that variation only

When generation quality is poor:

• Add more specificity to the subject description
• Strengthen the style directive with concrete references
• Increase negative prompt coverage
• Try switching between pro and flash models
• See references/prompt-patterns.md → Section 7 (Refinement Tips) for systematic troubleshooting

Step 5: Save + Log

Save images:

• Path: [project]/images/[platform]-[topic]-[YYYYMMDD]-v[N].png
• Example: images/instagram-feed-growth-mindset-20260224-v1.png
• Save the generation prompt alongside: images/prompt-log.md (append, don't overwrite)

Prompt log format:

## [YYYY-MM-DD] [Platform] — [Topic]
- Model: [pro/flash/auto]
- Aspect ratio: [ratio]
- Prompt: [full prompt text]
- Negative: [negative prompt]
- Result: [v1: kept / v2: kept / v3: discarded — reason]

This log enables prompt reuse and iterative refinement across sessions.

Update creative-log.md:

| [YYYY-MM-DD] | image-creator | [platform] [topic] × [N]images | [style notes] | nanobanana ([model]) |

Present to user: Show all generated images and ask which to keep, modify, or regenerate. If the user wants changes, return to Step 3 with adjusted prompt parameters.

Quality Checklist

Before delivering, verify:

• Brand colors from visual-guidelines appear in the generated images
• Aspect ratio matches the target platform specification
• Image style aligns with visual-guidelines direction (photo/illustration/abstract)
• The image visually communicates the post copy's core message
• No unintended text, watermarks, or visual artifacts
• 2–3 variations generated (giving the user choice)
• Prompt log saved for future reference and refinement
• creative-log.md updated with execution record

What This Skill Does NOT Do

• Write post copy → That's post-writer's job
• Extract visual guidelines → That's visual-extractor's job
• Edit or retouch existing images → Out of scope for v2
• Generate video or animation → Future extension (video-scripter)
• Write to brand-memory/ → Read-only access

Image Creator stays focused: copy in → brand-aligned images out.

Skill Chaining

← From post-writer: Each post includes a "Visual Direction" note — feed that directly into Step 1.

← From series-planner: Campaign plans include per-post visual direction — generate images for each post in sequence.

→ To user: Final deliverable — copy + matching visual, ready to publish.

Full pipeline: visual-extractor → framework-builder → trend-scout → post-writer → image-creator