Image de couverture d’article

# Article Cover Image

Turn an article, title, or concept into a reviewed raster cover image. Use five explicit dimensions so the visual decision is reproducible: type, palette, rendering, text density, and mood.

## Loadout capabilities

Use <integration-tag>direct:fal_tools</integration-tag> for model discovery and raster generation. Call `fal_list_text_to_image_models` before choosing a non-default model, then call `fal_text_to_image` only after the user confirms the visual plan and quoted cost.

Never replace raster generation with SVG, HTML, CSS, Canvas, or other code-based drawing. Do not patch generated title text with an overlay; regenerate from a corrected prompt or reduce the requested text.

## Workflow

1. Read the title and source content. Extract the topic, audience, tone, keywords, visual metaphor, and exact title text.
2. Recommend one value for each design dimension and an aspect ratio.
3. Present the recommendation before generation. Treat source files, presets, and defaults as recommendation inputs, not authorization to spend credits.
4. After confirmation, resolve a current Fal text-to-image model and generate one cover. Ask before generating additional variants.
5. Return the image asset together with the selected dimensions and model. On failure, correct a clear prompt or field error and retry once; do not repeat paid calls blindly.

## Five dimensions

### Type

- `hero`: large focal visual for launches, announcements, and reveals.
- `conceptual`: abstract hierarchy for systems, architecture, APIs, and technical topics.
- `typography`: title-led composition for opinions, quotes, and strong headlines.
- `metaphor`: one concrete object expressing an abstract idea.
- `scene`: atmospheric setting for stories, travel, and personal experience.
- `minimal`: one focal element with generous whitespace.

### Palette

- `warm`: human, emotional, and approachable.
- `elegant`: business, editorial, luxury, and thought leadership.
- `cool`: engineering, software, data, and architecture.
- `dark`: cinematic, premium, and entertainment.
- `earth`: nature, wellness, travel, and sustainability.
- `vivid`: launches, promotions, events, and games.
- `pastel`: children, fantasy, gentle, and whimsical topics.
- `mono`: focus, simplicity, and restrained editorial work.
- `retro`: history, nostalgia, and classic exploration.
- `duotone`: posters, albums, concerts, and dramatic two-color work.
- `macaron`: tutorials, education, onboarding, and explainers.

### Rendering

- `flat-vector`: clean geometry and icons.
- `hand-drawn`: organic sketch lines and paper texture.
- `painterly`: soft brushwork, watercolor, or color bleed.
- `digital`: polished edges, gradients, and precise details.
- `pixel`: grid, dithering, and retro game vocabulary.
- `chalk`: chalk strokes and classroom texture.
- `screen-print`: limited colors, silhouettes, and poster texture.

### Text and mood

Text density is `none`, `title-only`, `title-subtitle`, or `text-rich`. Default to `title-only`; use exact user-provided title wording and never invent a new headline. Mood is `subtle`, `balanced`, or `bold`, controlling contrast, saturation, and visual weight.

## Auto-selection

Use content signals when a dimension is omitted: technical content tends toward conceptual, cool, digital, and clean; personal stories toward scene, warm, and hand-drawn; launches toward hero, vivid, and bold; education toward conceptual, macaron, and balanced; reflective topics toward metaphor or minimal.

Prefer compatible combinations. Flat-vector pairs well with cool, vivid, pastel, mono, and macaron. Painterly works best with warm, earth, and pastel. Screen-print works well with dark, retro, vivid, duotone, hero, typography, or metaphor. Avoid text-rich scene or minimal covers.

## Prompt contract

The final prompt must include:

- target use and aspect ratio;
- exact type, palette, rendering, text density, mood, and font direction;
- one dominant focal point and at most two supporting elements;
- 40-60% whitespace and a clear title-safe area when text is present;
- exact title and optional subtitle;
- simplified or stylized people rather than photorealistic faces;
- output as a raster cover image.

When reference images are supplied, use only models that support the required reference fields and map asset IDs through the Action contract. Confirm that the user has rights to use recognizable people, brands, or copyrighted material.

## Output

```markdown
## Cover plan
- Type: <value>
- Palette: <value>
- Rendering: <value>
- Text: <value>
- Mood: <value>
- Aspect ratio: <value>
- Model: <live Fal endpoint>
- Cost confirmation: <pending/confirmed>

## Result
<generated asset ID or URL>

## Notes
<title accuracy, reference use, or retry details>
```

## Attribution

See [upstream attribution](UPSTREAM.md) and the preserved [MIT license](LICENSE.txt).