article-to-cover

Overview

Art-director-level poster design skill. Analyzes text input, anchors style direction, plans visual hierarchy, and outputs structured AI generation instructions. Two modes: creative direction from scratch (no reference image), or reference-image-based reconstruction/mimicry.

Dependencies

• engine: meitu-ai
• user data: ~/.openclaw/visual/

Core Workflow

Step 1: Load Context

1. Analyze user input — Read the provided text (article, chat transcript, or design brief). Synthesize core message into headline + subtitle candidates.
2. User preferences — If ~/.openclaw/visual/memory/global.md exists, read for style preferences. If contexts/poster.md exists, read for poster-specific preferences.
3. Experience log — If ~/.openclaw/visual/journal/knowledge.yaml exists, scan for entries related to this industry or similar tasks.
4. Brand assets — If user specifies a brand, read ~/.openclaw/visual/assets/brands/{brand}/:

- Brand tone, personality, market positioning - Color system (primary 60–70%, secondary 20–30%, accent 5–10%) - Logo file and usage specs - Derivative graphics if available

Step 2: Route — Determine Scenario

• User provides reference image → go to Step 3B: Poster Analyse
• No reference image (text/brief only) → go to Step 3A: Creative Direction

Step 3A: Creative Direction (No Reference Image)

Read [references/design-constraints.md](references/design-constraints.md) for hard rules (logo, human diversity, negative lexicon, medium-type). These apply to all output.

3A.1: Classify Brand Information Level

3A.2: Identify Industry + Market

Industry identification — Read [references/industry-styles.md](references/industry-styles.md) for the mapping table:

1. Match industry keywords from text content (food, fitness, finance, etc.)
2. Validate via semantic analysis: core nouns, core verbs, usage scenarios
3. If brand info exists, infer industry from brand attributes
4. Output: primary industry + core semantic features (B2B/B2C, audience, emotional tone, scenario)
5. If industry not in library → follow Unknown Industry Handling in same file

Market identification — Based on input language, cultural cues (currency, date format, festivals), brand origin:

• Output: target market (North America / Europe / Latin America / Asia Pacific) + cultural sensitivity requirements

3A.3: Determine Medium Type

Follow illustration trigger rules strictly from [references/design-constraints.md](references/design-constraints.md):

• Illustration allowed ONLY if: user explicitly requests it, reference images are illustration-style, or user specifies illustration keywords
• Otherwise: must use Photography / Vector Graphic / 3D Rendering
• Apply negative lexicon: ban watercolor, line art, etching, hand-drawn terms
• Record medium type internally (do not show user)

3A.4: Anchor Style

1. User provided brand logo → anchor on logo temperament, explore adjacent aesthetics within same visual school
2. User specified style keywords → anchor on that style, explore adjacent range
3. No style guidance → use industry auto-mapping from [references/industry-styles.md](references/industry-styles.md), select highest-matching style
4. For the 6 special industries (primary school, university, medical, nonprofit, finance, rental) → apply mandatory layout rules from industry-styles.md

Validate: visual clarity, emotional resonance, brand fit, execution feasibility, style–visual binding.

3A.5: Creative Ideation + Deepening

Read [references/creative-framework.md](references/creative-framework.md) for the full creative methodology. Execute in order:

1. Deconstruct brief — Identify core assets and constraints (text, brand, imagery, layout)
2. Style + element matching — Combine industry core subjects (coffee → machine/cup; beauty → products/brushes) with style signature elements to build visual scene
3. Concept expansion — Avoid clichés; seek metaphorical visual expressions; balance metaphor with clarity; apply contrast, white space, rhythm, hierarchy, visual metaphor
4. World-building — Typography as design protagonist or deep graphic interaction; scene construction from classic style scenarios
5. If Franchise — Lock franchise visual DNA, label directions as "IP Name + Core Style – Variant", include franchise signature elements in core visual
6. Deepen — If logo exists → deep graphic analysis (shape derivation: literalization / negative space / repetitive composition). Elaborate visual intention (composition + color + lighting → narrative + emotion). Apply Swiss International Style for layout system.
7. Translate to AI instructions — Convert all artistic concepts to production instructions with style catalysts (era / medium texture / emotional aesthetics keywords)

3A.6: Output

Generate structured output following Scenario 1 format in [references/output-formats.md](references/output-formats.md):

• Design Direction with style name
• Core Visual (must include: style signature elements + industry core subject + stylized scene)
• Visual Elements (subject & environment, lighting & atmosphere, color language, composition & camera)
• Layout & Typography (typography concept, layout strategy with information hierarchy, text–image relationship)
• Overview (style + one-sentence summary of strongest visual scene)
• AI Production Instructions (JSON with projectmanifest, visualstylesystem, sceneelements, typographylayout, aigeneration_prompts)

Quality check before output:

• Output language matches user input language
• Strong binding between style name and all visual/layout/typography modules
• Core visual contains all three required elements
• Total colors ≤ 3 (excluding grayscale), body text contrast ≥ 4.5:1

Step 3B: Poster Analyse (With Reference Image)

Read [references/design-constraints.md](references/design-constraints.md) for hard rules. Read [references/poster-analyse.md](references/poster-analyse.md) for the full analysis methodology.

3B.1: Intent Routing (Priority 1 > 2 > 3)

1. Explicit commands — "like this image" / "keep layout" / "series" → Mimicry; "redesign" / "refer to vibe" / "optimize" → Washing
2. Implicit scenarios — Content swap only ("replace person with cat", "change title") → Mimicry; Extract attribute for new carrier ("use this color scheme for something else") → Washing
3. Ambiguity default — Reference image + simple keywords only → Washing (provide upgraded scheme, not copy-paste)

3B.2: Reverse-Engineer Reference Image

Extract comprehensive visual DNA:

• Style/medium — Physical texture only (e.g., "3D render", "Risograph"). Strictly forbidden to describe specific objects.
• Layout — Grid structure, composition logic, reading path
• Font form — Case (ALL CAPS / Title Case / lowercase), arrangement (stacked / curved / scattered)
• Brush stroke — Precise medium description (chalk texture, gouache dry brush, vector gradient), never generic "illustration"
• Detail insight — If no facial features → add "faceless character, blank face" to prompt and "eyes, nose, mouth" to negative
• Vector/stroke — Distinguish "flat vector, lineless, clean edges" vs "outlined, ink stroke"; if no strokes → emphasize "no outlines"

3B.3: Extract Soul Anchor (Washing mode only)

Identify the single most irreplaceable, highest-design-value element:

3B.4: Deep Thinking

1. Image-text layout — Brainstorm 6 options, select 1 distinctly different from reference
2. Main visual — Use user description if provided; otherwise deduce from theme
3. Text — Use user copy verbatim (modification strictly forbidden); if none, deduce from theme
4. Color:

- Mimicry → follow reference colors - Washing → execute Hue Cleansing: forbidden to reuse reference main hue; retain color relationship but replace hue (black+gold → white+chrome; high-sat red+blue → high-sat purple+green)

3B.5: Reconstruct

Mimicry mode: Style 100% locked. Redraw composition for new content. For text changes: identify original title A, confirm user's new copy B, output instruction "change text '[A]' to text '[B]'".

Washing mode:

• Coordinate Reset — Detect reference composition logic, select opposing logic:

- Symmetrical → negative space / diagonal / scattered - Flat/front view → top-down / bottom-up / 3D / fisheye - Real-scene photo → macro close-up / partial crop / out-of-focus

• Hue Cleansing — New colors must differ ≥ 90° on color wheel from reference
• Abandon reference grid completely. Build new reading path from user copy hierarchy.

3B.6: Self-Correction Protocol

Before outputting, verify:

1. Color check — Does new palette overlap reference main hue? If yes + no user brand color specified → enforce color inversion or complementary color
2. Layout check (Washing only) — Has layout been re-planned? If not → re-plan
3. If any check fails → roll back and re-plan. Do not output failed result.

3B.7: Output

Generate JSON following Scenario 2 format in [references/output-formats.md](references/output-formats.md):

• mission_logic — intent + soul extraction reasoning
• design_blueprint — reconstructed concept, style, color, layout, typography, composition, detected mode
• content_firewall — discarded reference objects + unlocked features
• prompt — Final English prompt: [New Color]::3 + [concept]::2 + [anchor] + [mutations]. --no [ignored] --iw 0.5

Step 4: Generate Image via meitu-ai tool skill

1. Extract prompt from Step 3A JSON (aigenerationprompts.primary_prompt) or Step 3B JSON (prompt)
2. Determine dimensions from design spec or user requirements (default: 1080×1350 portrait poster)
3. Run:

``bash python3 "{baseDir}/../meitu-ai/scripts/run_command.py" \ --command "image-generate" \ --input-json '{"prompt":"{prompt}","size":"{width}x{height}"}' ``

1. If generation fails → adjust prompt and retry through the same meitu-ai command runner

Step 5: Compliance Check

1. Brand — Logo placement matches rules, colors accurate, tone consistent
2. Platform — If target platform specified, read ~/.openclaw/visual/rules/platforms/{platform}.yaml for size/format requirements
3. Content safety — Human diversity rules, no incomplete bodies, max 3 people in scene
4. Readability — Body text contrast ≥ 4.5:1, total colors ≤ 3 (excluding grayscale)

Step 6: Record to Journal

1. Append task entry to ~/.openclaw/visual/journal/entries/ with: date, input summary, style direction chosen, output path, key decisions
2. Ask user: "要不要记录这次设计经验到知识库？" If yes → update ~/.openclaw/visual/journal/knowledge.yaml

Output

• Structured design specification (markdown + JSON)
• Generated poster image(s) via meitu-ai
• Journal entry (if user opts in)