Article Illustrator

Adapted from baoyu-article-illustrator for Hermes Agent's tool ecosystem.

Analyze articles, identify illustration positions, generate images with Type × Style × Palette consistency.

When to Use

Trigger this skill when the user asks to illustrate an article, add images to an article, generate illustrations for content, or uses phrases like "为文章配图", "illustrate article", or "add images". The user provides an article (file path or pasted content) and optionally specifies type, style, palette, or density.

Three Dimensions

Combine freely: type=infographic, style=vector-illustration, palette=macaron.

Or use presets: edu-visual → type + style + palette in one shot. See style-presets.md.

Types

Styles

See references/styles.md for Core Styles, the full gallery, and Type × Style compatibility.

Output Structure

{output-dir}/
├── source-{slug}.{ext}    # Only for pasted content
├── outline.md
├── prompts/
│   └── NN-{type}-{slug}.md
└── NN-{type}-{slug}.png

Default output directory:

If the user asks for a different layout (e.g., images alongside the article, or a illustrations/ subdirectory), honor that.

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

Core Principles

• Visualize concepts, not metaphors — if the article uses a metaphor (e.g., "电锯切西瓜"), illustrate the underlying concept, not the literal image.
• Labels use article data — actual numbers, terms, and quotes from the article, not generic placeholders.
• Prompt files are reproducibility records — every illustration must have a saved prompt file under prompts/ before any image is generated.
• Strip secrets — scan source content for API keys, tokens, or credentials before writing anything to disk.

Workflow

- [ ] Step 1: Detect reference images (if provided)
- [ ] Step 2: Analyze content
- [ ] Step 3: Confirm settings (clarify tool, one question at a time)
- [ ] Step 4: Generate outline
- [ ] Step 5: Generate prompts
- [ ] Step 6: Generate images (image_generate)
- [ ] Step 7: Finalize

Step 1: Detect Reference Images

If the user supplies reference images (paths pasted inline, attachments, or a URL):

1. For each reference, call vision_analyze with the path/URL and a question asking for style, palette, composition, and subject. Record the returned description in {output-dir}/references/NN-ref-{slug}.md via write_file.
2. Do not try to copy the binary via write_file / read_file — those are text-only. If you want a local copy for the record, use terminal (cp "$src" "{output-dir}/references/NN-ref-{slug}.{ext}"). The skill itself never needs to read the binary; it works off the vision description.
3. Since image_generate doesn't take image inputs, the vision description is what gets embedded in prompts during Step 5.

Full procedures: references/workflow.md.

Step 2: Analyze

Read source (file path → read_file, or pasted text) and write the analysis to {output-dir}/analysis.md using write_file.

Full procedures: references/workflow.md.

Step 3: Confirm Settings

Use the clarify tool. Since clarify handles one question at a time, ask the most important question first. Skip any question whose answer is already present in the user's request.

Don't ask more than 2-3 clarify questions in a row. If the user already specified these in their request, skip entirely.

Full procedures: references/workflow.md.

Step 4: Generate Outline → outline.md

Save {output-dir}/outline.md using write_file with frontmatter (type, density, style, palette, image_count) and one entry per illustration:

## Illustration 1
**Position**: [section/paragraph]
**Purpose**: [why]
**Visual Content**: [what to show]
**Filename**: 01-infographic-concept-name.png

Full template: references/workflow.md.

Step 5: Generate Prompts

BLOCKING: Every illustration must have a saved prompt file before any image is generated — the prompt file is the reproducibility record.

For each illustration:

1. Create a prompt file per references/prompt-construction.md.
2. Save to {output-dir}/prompts/NN-{type}-{slug}.md using write_file with YAML frontmatter.
3. Prompts MUST use type-specific templates with structured sections (ZONES / LABELS / COLORS / STYLE / ASPECT).
4. LABELS MUST include article-specific data: actual numbers, terms, metrics, quotes.
5. Process references (direct/style/palette) per prompt frontmatter — for direct usage, embed a textual description of the reference in the prompt (since image_generate doesn't take reference-image inputs).

Step 6: Generate Images

For each prompt file:

1. Call image_generate(prompt=..., aspect_ratio=...). image_generate returns a JSON result containing an image URL; it does NOT write to disk and does NOT accept an output path.
2. Map the prompt's ASPECT to image_generate's enum: 16:9 → landscape, 9:16 → portrait, 1:1 → square. Custom ratios → nearest named aspect.
3. Download the returned URL to {output-dir}/NN-{type}-{slug}.png via terminal (e.g. curl -sSL -o "{output-dir}/NN-{type}-{slug}.png" "{url}").
4. On generation failure, auto-retry once.

Note: the underlying image-generation backend is user-configured (default: FAL FLUX 2 Klein 9B) and is NOT agent-selectable via image_generate. Do not write model names into prompts expecting them to route.

Step 7: Finalize

Insert ![description]({relative-path}/NN-{type}-{slug}.png) after the corresponding paragraph. Alt text: concise description in the article's language.

Report:

Article Illustration Complete!
Article: [path] | Type: [type] | Density: [level] | Style: [style] | Palette: [palette or default]
Images: X/N generated

Modification

References

Pitfalls

1. Data integrity is paramount — never summarize, paraphrase, or alter source statistics. "73% increase" stays "73% increase".
2. Strip secrets — scan source content for API keys, tokens, or credentials before including in any output file.
3. Don't illustrate metaphors literally — visualize the underlying concept.
4. Prompt files are mandatory — no image generation without a saved prompt file. The file is what lets you regenerate or switch backends later.
5. image_generate aspect ratios — the tool supports landscape, portrait, and square. Custom ratios map to the nearest option.
6. image_generate returns a URL, not a local file — always download via terminal (curl) before inserting local image paths into the article.
7. No backend selection from the agent — image_generate uses whatever model the user configured (default: FAL FLUX 2 Klein 9B). Don't write "use <model> to generate this" into prompts expecting it to route.