I built an AI agent skill that turns article concepts into weird, hand-drawn illustrations

#ai #showdev #opensource #writing

Every technical article eventually hits the same wall: you've written something you're proud of, and now you need images. So you reach for the usual suspects — a stock photo of someone pointing at a whiteboard, a generic gradient banner, maybe a flowchart that looks like every other flowchart on the internet. Nobody remembers any of it.

I wanted the opposite of that. I wanted illustrations that actually say something — that take a judgment or a process buried in a paragraph and turn it into one small, strange, memorable picture.

So I built Kunaal's Illustrations — an AI agent skill that does exactly this.

What it actually is

It's not a prompt. It's not a PPT template. It's a skill you drop into your AI agent that teaches it a very specific way of thinking about illustrations.

Instead of "add a picture here," the agent learns to ask: what's the one cognitive action in this paragraph worth drawing? Then it draws that — in a consistent hand-drawn style, with a recurring cast of characters who are genuinely part of the scene rather than decoration standing in the corner.

Here's the boy saying hi. He shows up a lot.

The core idea: draw the concept, not decoration

This is the whole philosophy in one line: the character has to do the work.

If you could remove the character from the picture and the picture would still make sense, then the character was just decoration — and that's a fail. Every illustration has to hang on a single idea, and a person (or occasionally a very unbothered cat) has to physically act it out.

A few examples that came out of it:

"Too many tasks, zero focus." A boy hopping between floating paper islands, exhausted, never actually landing on any of them.

"Messy thought becomes structure." A girl at her desk with a tangled cloud of thoughts above her head, and clean structured lines forming on the page under her pen.

"Preparation, meet reality." A boy throwing his papers in the air because his grades didn't match the effort.

None of these needed a caption to land. That's the point.

The visual DNA

The style is deliberately restrained so everything looks like it came from the same hand:

Pure white background — no paper texture, no shadows, no gradients.
Black hand-drawn line art with thin, slightly wobbly lines.
Lots of white space — the subject only fills about 40–60% of the frame.
Sparse red / orange / blue handwritten labels, and only when a label genuinely helps.
One image = one idea. No cramming a whole chapter into a single picture.

The two default characters are a boy (short black hair, light blue t-shirt, grey pants) and a girl (shoulder-length black hair, yellow t-shirt, dark grey pants). They're not mascots — they're the workers doing the absurd job the concept requires.

Who it's for (and who it isn't)

It's a great fit if you:

write English articles, blogs, or Notion docs and want in-text illustrations
create methodology, knowledge, or AI-workflow content
like turning abstract judgments into concrete, slightly weird metaphors
want a reusable visual language that stays consistent across everything you publish

It's not for you if you want polished brand key visuals, corporate flat illustrations, dense architecture diagrams, cutesy emoji-style mascots, or strictly editable vector source files. That's a different tool.

How it works under the hood

The skill runs a small, opinionated workflow rather than firing off a single image:

Digest the text — read the article and find the cognitive anchors: the core judgments, the turning points, the before/after moments.
Plan first, draw later — output a shot list of 4–8 images. For each one it writes the paragraph it follows, the theme, the core meaning, the structure type, what the character is doing, and suggested labels.
Generate one image at a time — never stitched together, each carrying exactly one idea.
QA against a checklist — is the background clean white? Is there enough white space? Is the character actually doing something? Does it accidentally look like a PowerPoint slide? If it fails, regenerate.
Save and report — final PNGs land in a Generated/ folder, with a note on which images are solid and which are optional.

You can also stop it at step 2 if you just want the illustration plan and want to draw things yourself.

Trying it out

The fastest path (the way I use it):

Open Antigravity and select the Gemini 3.1 Pro (High) model.
Paste the repo's .git URL into the chat and tell the agent to clone/install it.
Start prompting.

Or install it manually into your agent's skills directory:

git clone https://github.com/Yuvakunaal/kunaal-illustrations.git
cd kunaal-illustrations
cp -R ./kunaals-illustrations ~/.gemini/config/skills/

Then the prompts are refreshingly simple. Whole-article mode:

Use $kunaals-illustrations to generate 4 bizarre in-text illustrations for the article below.
Requirements: wide composition, pure white background, black hand-drawn line art,
sparse red/orange/blue English handwritten annotations.

<paste your article>

Planning-only mode:

Use $kunaals-illustrations, do not generate images yet.
Analyze where this article is worth illustrating and output a shot list of about 5 images.

<paste your article>

Single concept:

Use $kunaals-illustrations to generate an in-text illustration for
"Trust is not shouted out, but built piece by piece with evidence."
Make it bizarre but clean, and the boy must carry the core action.

A couple of things I've learned using it: keep the label text short (the fewer words in the image, the more stable the output), and let each picture stay about one idea. AI image models will still occasionally hallucinate a typo or drift on style — when that happens, cutting the number of labels and regenerating usually fixes it.

A note on licensing

The repo is dual-licensed: the skill code, prompt templates, and docs are MIT, and the example images, character IPs, and style references are CC BY 4.0. If you use or adapt the characters or style, please credit "Kunaal's Illustrations" by Kunaal and link back to the repo.