| name | banana-paper-illustration |
| description | Generate paper-style illustrations with the Banana/Jiekou text-to-image API when the user wants a visual abstract, teaser figure, concept illustration, method overview art, or publication-style research image from a paper abstract, method description, caption, or technical note. Use this skill for image-first scientific figures, not for precise charts, tables, or editable box-arrow diagrams. |
Banana Paper Illustration
Generate publication-style research illustrations with the Banana text-to-image API.
Use this skill when the user wants:
- a paper teaser image
- a visual abstract
- a polished concept illustration for a method or system
- a paper figure that should look more visual than a draw.io block diagram
Do not use this skill when the user needs:
- exact plots or charts
- precise tables
- patent black-and-white line drawings
- editable architecture diagrams with strict node/edge control
For editable structure-first diagrams, prefer drawio-architecture-diagram.
Workflow
1. Decide Whether Image Generation Is Appropriate
Use this skill only if the target figure is image-first.
Good fits:
- visual summaries
- conceptual figures
- research covers
- system atmosphere figures
- hybrid method overview illustrations with limited labels
Bad fits:
- algorithm flowcharts with exact topology
- benchmark plots
- diagrams requiring deterministic box placement
2. Build The Figure Intent
Extract from the user's source:
- the paper or method title
- the central scientific story
- 3-6 visual elements
- what must be foreground vs background
- whether arrows, stages, agents, or feedback loops should appear
- what style is expected
If the source is long, compress it before prompting. Do not dump raw sections into the API.
3. Choose A Figure Mode
Use one of these modes:
teaser: visually strong paper opener with one dominant composition
method-overview: a semi-structured method illustration with a few modules or arrows
visual-abstract: balanced high-level summary for the whole paper
system-concept: system or architecture concept art with technical atmosphere
4. Generate A Prompt
Use the bundled prompt logic in:
Read these references as needed:
The script can:
- accept a direct prompt
- derive a prompt from a source file
- auto-select a palette by domain when
--palette is not specified
- apply a scientific palette preset such as
vivid-academic or okabe-ito
- save the final resolved prompt to a sibling
.prompt.txt file
- call the API and write the returned image to disk
5. Call The API
Use the script instead of hand-writing the request each time:
python3 scripts/generate_banana_illustration.py \
--source-file paper_method.md \
--mode method-overview \
--palette vivid-academic \
--output paper_figure.png
The script reads the API key from:
BANANA_API_KEY
API_KEY
Do not hard-code secrets into the skill or output files.
6. Review And Iterate
After generation, review:
- whether the scientific story is clear
- whether the composition matches the intended figure type
- whether the image uses too much unreadable micro-text
- whether key modules or agents are visually distinct
- whether the output looks like a paper figure instead of generic concept art
If the result is too loose:
- simplify the figure story
- reduce the number of modules
- add stronger composition instructions
- switch to
drawio-architecture-diagram if exact structure matters more than visual richness
Output Contract
By default, produce:
1. Figure Intent
State:
- chosen figure mode
- why image generation fits better than draw.io for this request
2. Prompt Summary
Return:
- a short summary of the final prompt direction
- the main scientific elements being visualized
- the key style constraints
3. Generated Files
Return:
- image path
- prompt file path
- source file path if one was created
4. Risks
Call out the main failure risks, such as:
- unreadable text inside the image
- structure drift
- too much artistic freedom
- weak faithfulness to the paper
Practical Rules
- Prefer light or white backgrounds unless the user explicitly wants a dark poster style.
- Prefer sparse labels; text rendering inside images is fragile.
- Keep the scientific story narrow. One figure should express one main idea.
- For method illustrations, ask for a structured composition but do not expect pixel-perfect topology.
- If the user needs strict module placement, switch to draw.io instead of fighting the image model.
Validation
Before finalizing, verify:
- the prompt is specific enough to encode the main scientific story
- the requested mode matches the actual figure need
- the output path is explicit
- the API key is read from environment variables only
Failure Handling
If the API key is missing:
- stop with a clear error naming the supported environment variables
If the API returns non-image content:
- save or print the diagnostic response
- do not pretend the image succeeded
If the result is too diagram-like but still inaccurate:
- recommend switching to
drawio-architecture-diagram