Baoyu Cover Image

Cover Image Generator

Adapted from baoyu-cover-image for Hermes Agent.

Generate elegant article cover images with 5-dimensional customization: Type × Palette × Rendering × Text × Mood. Supports cinematic (2.35:1), widescreen (16:9), square (1:1), and custom aspect ratios.

When to Use

Trigger when the user asks to "generate cover image", "create article cover", "make cover", "封面图", or provides an article and wants a cover. The user provides content (file, URL, or pasted text) and optionally specifies dimensions.

Five Dimensions

Type (Visual Composition)

Palette (Color Scheme)

11 palettes: warm (personal/story), elegant (business/luxury), cool (tech/system), dark (cinematic/premium), earth (nature/wellness), vivid (launch/promotion), pastel (fantasy/gentle), mono (zen/focus), retro (history/vintage), duotone (movie poster/dramatic), macaron (education/tutorial).

Rendering (Visual Style)

Text Level

Mood (Emotional Intensity)

Style Presets (shortcuts)

--style <name> sets Palette + Rendering in one shot. Explicit --palette/--rendering overrides preset values.

Auto-Selection

When dimensions are omitted, select based on content signals:

Type

Palette

Rendering

Text

Mood

Font

Output Structure

cover-image/{topic-slug}/
├── source-{slug}.md       # Saved source content
├── prompts/cover.md       # Generation prompt (reproducibility record)
└── cover.png              # Output image

Slug: 2-4 words kebab-case. Conflict: append -YYYYMMDD-HHMMSS.

Workflow

- [ ] Step 1: Analyze content
- [ ] Step 2: Confirm options (clarify)
- [ ] Step 3: Create prompt → prompts/cover.md
- [ ] Step 4: Generate image → cover.png
- [ ] Step 5: Completion report

Step 1: Analyze Content

1. Save source content (file → read_file, pasted text → write_file to source-{slug}.md)
2. Analyze: topic, tone, keywords, visual metaphors
3. Detect language (source, user input, or explicit --lang)
4. Determine output directory: cover-image/{topic-slug}/
5. Strip secrets: scan for API keys, tokens, credentials — never include in outputs

Step 2: Confirm Options

Use the clarify tool. Ask at most 2-3 questions in sequence:

Q1 — Type, Palette, Rendering: Present 2-3 auto-recommended combinations with rationale. User picks one. Q2 — Text & Aspect: Confirm text level and aspect ratio (16:9 default, or 2.35:1/4:3/3:2/1:1). Q3 — Language (only if source ≠ user language): which language for title text?

Skip questions where user already specified values. Skip entirely if user said "quick" or "不用确认".

Step 3: Create Prompt → prompts/cover.md

Write a structured prompt file with YAML frontmatter:

---
type: cover
palette: [confirmed]
rendering: [confirmed]
---

# Content Context
Article title: [exact title from source]
Content summary: [2-3 sentences]
Keywords: [5-8 terms]

# Visual Design
Type: [confirmed] | Palette: [confirmed] | Rendering: [confirmed]
Font: [confirmed] | Text level: [confirmed] | Mood: [confirmed]
Aspect ratio: [confirmed] | Language: [confirmed]

# Text Elements
[Title/subtitle/tags per text level — use exact title from source; never invent]

# Composition
[Type-specific layout + visual metaphor from content + palette colors + rendering notes]

Principles:

• Title text: Use exact title from source/user. Do NOT invent.
• Visual metaphor: Derived from content meaning, not literal illustration
• Whitespace: 40-60% breathing room
• Characters: Simplified silhouettes; NO realistic humans
• Color constraint: hex values are rendering guidance — do NOT display as visible text in image
• Strip secrets from all output files

Step 4: Generate Image

1. Write full prompt to prompts/cover.md (reproducibility record — MUST do before generation)
2. Call image_generate(prompt=..., aspect_ratio=...) with the prompt content
   • Map aspect: 16:9/4:3→landscape, 9:16/3:4→portrait, 1:1→square
3. image_generate returns a URL — download via terminal:

   curl -sSL -o "/absolute/path/to/cover-image/{slug}/cover.png" "{url}"
   

4. Verify file exists and is non-empty
5. On failure, auto-retry once

Prompt file requirement (hard): write prompts/cover.md BEFORE calling image_generate. The file is the reproducibility record.

Step 5: Completion Report

Cover Generated!

Topic: [topic]
Type: [type] | Palette: [palette] | Rendering: [rendering]
Text: [text] | Mood: [mood] | Font: [font] | Aspect: [ratio]
Language: [lang]

Files:
✓ source-{slug}.md
✓ prompts/cover.md
✓ cover.png

Composition Principles

• Whitespace: 40-60% breathing room; avoid clutter
• Visual anchor: Main element centered or offset left (right side for title)
• Characters: Simplified silhouettes only — NO realistic human faces
• Title: Use exact title from user/source; never invent
• Icon vocabulary: Simple recognizable icons (code window, lightbulb, gear, lock, etc.) rather than detailed illustrations

Modification

Pitfalls

1. Data integrity: Use exact title from source — never invent or modify
2. Strip secrets: Scan for API keys, tokens, credentials before writing any file
3. Prompt file first: Write prompts/cover.md BEFORE image_generate — it's the reproducibility record
4. image_generate returns URL not file: Always download via curl -sSL -o {absolute path} {url}
5. Aspect ratio mapping: 16:9→landscape, 9:16→portrait, 1:1→square; custom ratios→nearest
6. No backend selection from agent: image_generate uses user-configured backend; don't write model names into prompts
7. Visualize concepts not metaphors: If article says "电锯切西瓜", illustrate the underlying concept, not the literal image
8. No text repair on bitmaps: If title is garbled, regenerate from corrected prompt — never overlay/Photoshop