AI Whiteboard Doodle Cards
One topic, one deck
Turn any topic into a set of black-marker stick-figure doodle cards. The English track lets the model bake short labels right into the art; the Chinese track draws textless bases and overlays handwriting with PIL — locked style, so the whole set looks drawn in one sitting.
10
cards / set
EN+ZH
two tracks
4
credits/img (low)
5
pipeline stages
What you can do after this
- Break a topic into 8–12 one-fact-per-card doodle beats
- Keep a whole set visually consistent with a locked style (RSA-animate whiteboard look)
- English track: have gpt-image-2 bake short labels into the doodle
- Chinese track: generate textless bases, then overlay clean handwriting with PIL (the model can't spell Chinese)
- Drive everything from cards.json — edit copy without redrawing
- Lock facts with multi-source checks so model-hallucinated stats never ship
Course materials
Submit to downloadShare a few details and your topic of interest to download the sample deck, caption template, and skill spec.
To share materials with interested learners, leave your basic info and topics first. The file opens automatically after submission.
Course structure
From locking facts, to designing cards, to two generation tracks, to QA and publishing.
Lock the facts
Multi-source verification, one fact per line with sources
Design cards.json
Narrative arc, one fact per card, short labels
English track
Model bakes text, prompt locks the style
Chinese track
Textless base + PIL handwriting overlay
Compose & QA
Pick large / overlay, contact sheet, check for garbled text
Caption & publish
Social caption + ship to nextagent.ca
Bake English, overlay Chinese
Image models spell short English but not Chinese — so the two tracks split, and a locked style keeps the whole set coherent.
The standard loop
- Research: pull 2+ authoritative sources and nail every number
- Design: one fact per card in cards.json
- English: make_prompts --lang en → gen → finalize (pick large)
- Chinese: make_prompts --lang cn → gen (textless) → finalize (overlay)
- QA: read the contact sheet, regenerate only the garbled card
- Caption: title / body / sources / disclaimer
Three pillars
Lock the facts
Multi-source
Every record on one line with a source; events past the knowledge cutoff must be verified online — never invent a result.
Data-driven
cards.json
One JSON holds every card; the scripts take the English head or Chinese textless head by --lang. Edit copy without redrawing.
Two tracks
Bake vs overlay
English is baked into the image by the model; Chinese is a textless base with handwriting overlaid via PIL.
Pre-publish checklist
Before a deck ships, confirm at least: 1. Does every card map to a verified fact? 2. Any model-hallucinated stats baked in (e.g. "possession 62%")? 3. Are Chinese names in the caption, not forced into the base image? 4. Is any headline text garbled? 5. Are all source links present in the caption? 6. Is it marked "AI-generated sample; data belongs to original sources"?
Suggested teaching pace
| Time | Duration | Module | Mode |
|---|---|---|---|
| 00:00–10:00 | 10 min | Why one fact per card | Talk |
| 10:00–25:00 | 15 min | Lock facts + design cards.json | Break down |
| 25:00–45:00 | 20 min | English track: bake-text prompts + generate | Hands-on |
| 45:00–65:00 | 20 min | Chinese track: textless base + overlay | Hands-on |
| 65:00–80:00 | 15 min | QA + social caption + publish | Discuss |
Who it's for
For anyone who needs to turn a topic into a full set of visual content fast.
A fit ✓
- Creators / social media operators
- AI content-production consultants
- Teams making explainer carousels
- Claude Code / OpenClaw users
Not a fit ✗
- Need photoreal illustration
- Won't verify facts before posting
- Want a single poster, not a set (see mmc-longposter)
FAQ
Why does Chinese need its own track?
gpt-image-2 can't spell Chinese — it comes out garbled. So Chinese cards are generated as digit-only textless bases, then clean handwritten title and caption are overlaid with PIL.
Do I re-generate images to change copy?
Not for Chinese — the text is an overlay. Edit cn_title / cn_caption in cards.json and re-run finalize: instant, zero credits. English text is baked in, so changing it means regenerating that card.
What does it cost?
Default low quality is 4 credits/image — about 40 for a 10-card set, visually plenty for doodles.
Can I trust the numbers on the cards?
Headline facts are multi-source verified; but the image model sprinkles in decorative micro-text (e.g. "possession 62%") that is NOT real data. Real facts live in the caption and the cards.json fact field — verify before publishing.
Is it open source?
The tooling and full bilingual example are open at github.com/NextAgentBC/doodle-cards.
Want your topic turned into a deck?
We can wire research, design, generation, and captions into one reusable visual-content line for you.