Core positioning
Design and generate 16:9 horizontal body illustrations for Chinese articles. The goal is not to make commercial illustrations, PPT infographics or cute cartoons, but to turn the key judgments, processes, structures, states or metaphors in the article into a clean, grotesque, creative, readable but non-instructional hand-drawn explanatory image.
The default visual IP is "Xiaohei": solid black, white dot eyes, thin legs, blank expression, seriously doing one absurd but valid thing. Xiaohei must participate in the core action of the image, and cannot just stand nearby as decoration.
First read these references
Read as needed for the task, do not stuff the context all at once:
references/style-dna.md: style DNA, colors, text, taboos.references/xiaohei-ip.md: the image, personality, action library and taboos of the Xiaohei IP.references/composition-patterns.md: structure types, original metaphor methods and anti-repetition rules.references/prompt-template.md: single-image generation prompt template.references/qa-checklist.md: post-generation checking and iteration rules.assets/examples/: only for low-frequency visual calibration, not part of the default generation path. Do not copy the composition, objects or annotations of these cases.
Workflow
1. Digest the body text
First read the body text, link, Notion page, Markdown file or screenshot content provided by the user. Extract:
- What the core viewpoint is
- Which paragraphs carry the cognitive turning points
- Which content is suitable to be explained with images
- Which places are only suitable for text and do not need images
Do not distribute illustrations evenly. Prioritize "cognitive anchors", for example: core judgments, two breakpoints, input-output closed loops, branching, before-and-after comparison, one fish many ways, handoff paths, common pitfalls, character state changes.
2. First produce the illustration strategy
If the user only says "analyze how to illustrate / think about which places need illustrations", first give a shot list. For each image, write clearly:
- After which paragraph it is placed
- The theme of the image
- The core meaning
- Structure type
- What Xiaohei does in the image
- Suggested elements
- Suggested Chinese annotation words
Default 4-8 images. When the article is very short, 1-3 images; for long articles, also do not easily exceed 9 images. Enough is enough, avoid turning the body text into an art book.
3. Single-image generation
If the user explicitly requests "generate / output / make image / help me generate", do not stop to wait for confirmation; use the built-in image_gen Generate each image separately. Do not combine multiple images into one.
Each image should explain only one core structure. The prompt must include:
- 16:9 horizontal Chinese body illustration
- Pure white background
- Black hand-drawn line art
- A small amount of red/orange/blue Chinese handwritten annotations
- Lots of white space
- Xiaohei as the subject of the core action
- Prohibit PPT, commercial illustration, childish cuteness, complex architecture, type title in the upper-left corner
Do not replicate past cases. Cases only provide style density and how Xiaohei participates; do not directly reuse existing compositions such as "conveyor belt breakpoint / Xiaohei pulling a line / asset fish / stamp toolbox / common pitfalls path" unless the user explicitly asks to replicate a certain image. Each time, reinvent a strange but valid metaphor from the current article.
4. Check and iterate
After generation, check references/qa-checklist.md. If any of the following issues appear, prioritize regenerating or editing locally:
- Xiaohei is merely decorative
- The image is too crowded
- It looks too much like a flowchart/PPT
- Too much Chinese or serious typos
- Titles such as "common pitfalls/flowchart/system architecture diagram" appear in the upper left corner
- The style is too cute, childish, or rigid
- The background is not a clean white background
5. Save and deliver
If the user is working inside the workspace, copy the final image to:
assets/<article-slug>-illustrations/
Name them in order:
01-topic-name.png
02-topic-name.png
Keep the original generated files; do not overwrite existing assets unless the user explicitly asks to replace them.
Output specifications
The pre-generation strategy output should be short and accurate. The post-generation delivery should include:
- How many images were generated
- The purpose of each image
- Save paths
- Which images are the most reliable and which are optional
Do not give long explanations of style theory; let the images speak for themselves.