Verdant Skyline

I made an ink sketch you can throw things at, and I want to talk about how it got made.

It started with a reference: a blue ballpoint landscape someone had drawn into a website footer. My first prompt asked for "a simple webpage with a canvas that contains dynamic drawn pen strokes like the image" and ended with "Let's brainstorm."

Sixteen prompts and two sittings later, it's a garden city in one-point perspective. The ink scatters around your cursor, walkers wander the terrace, and a click lobs a ball that blasts holes in walls, snaps tree limbs and knocks people over. They get back up.

What changed is the distance between an idea and a running sketch. "The lines should break apart and avoid the mouse cursor" became working code in one turn.

What didn't change is the loop. Look at it, react, push. My prompts were ideas and reactions, not specs: "Push harder on the details." "The tree tops are disconnected from the trunk." Taste still drives.

Then the Comfy part. I work at Comfy, so I wanted to see what happens when the recording becomes the input. We sent the same five seconds, the tree coming down, to five video models on Comfy Cloud with the same brief: make it a watercolor and ink world and keep every motion. A blind panel of three AI judges scored them on motion, look and stability. Gemini Omni was the prettiest but rewrote the city. Wan 2.2, an open model, kept the layout best. Kling O3 was the steadiest, so it got the full shot.

On who did what: I brought the reference, the reactions and the calls on what came next. Claude Code wrote the code, tested it in a browser between turns, ran six rounds of adversarial review on itself, and ran the Comfy bake-off and the judging.

The video is every version, v1 to v15, rebuilt from the session log, then the Comfy ending. The sketch is plain JavaScript with no libraries, under 200 KB. Links are in the first comment.

I'm curious how others are running this loop right now. Which part do you keep for yourself?

First comment

Try it: https://thedatalife.github.io/sketch-ball/ (a mouse is best, but touch works too: tap to lob a ball, drag to scatter the ink)
Code: https://github.com/thedatalife/sketch-ball
The Comfy part ran on Comfy Cloud with Kling O3, Runway Aleph2, Seedance 2.5, Gemini Omni 1.1 Flash and Wan 2.2 Fun Control. Here's the whole bake-off as one workflow you can open and run: https://cloud.comfy.org/?share=59c5ee8251f8
Times in the video are Pacific. The two sittings were Sep 22 and Sep 23, and the Comfy run was Oct 16.

I made an ink sketch you can throw things at. 16 prompts, 15 versions, two sittings, and then a Comfy ending.

Here's every version, rebuilt from the session log, with the prompt that made it. 🧵

It started with a reference: a blue ballpoint landscape someone drew into a website footer.

My first prompt asked for "a simple webpage with a canvas that contains dynamic drawn pen strokes like the image" and ended with "Let's brainstorm."

A lot of my prompts were just reactions.

"This needs to have a significant amount more detail."
"Push harder on the details."
"Push the design even harder."

Look at the last version, nudge, repeat.

Some were physics in plain words: "The lines should break apart and avoid the mouse cursor. They should move like a vector field..."

That went from a sentence to a running sketch in one turn.

Some were math. I asked how to make the reveal land like paint splashes, mathematically. It came back with options and the equations behind them, and I picked burst + wobble over its recommendation.

Some were just looking closely: "The tree tops are disconnected from the trunk."

Trees took three tries. Floating tops, then bubble wrap, then branches grown toward the crown.

Then it got pretty silly. A click lobs a ball from the bottom of the view. It blasts holes in walls, snaps tree limbs and knocks walkers over.

They get back up 🙂

Then the Comfy part. I work at Comfy, so I wanted the recording to become the input.

Same 5 seconds, same brief (make it watercolor and ink, keep every motion), five video models on Comfy Cloud in one batch.

Workflow: https://cloud.comfy.org/?share=59c5ee8251f8

A blind panel of three AI judges scored motion, look and stability.

Gemini Omni was the prettiest but rewrote the city. Wan 2.2, an open model, kept the layout best. Kling O3 was the steadiest, so it got the full 10 s shot.

Who did what: I brought the reference, the reactions and the calls on what came next. Claude Code wrote the code, tested it in a browser, ran six rounds of adversarial review on its own work, and ran the Comfy bake-off.

What changed: the distance from idea to running sketch, and from sketch to a new medium. What didn't: the loop. Look, react, push.

Try it: https://thedatalife.github.io/sketch-ball/
Code: https://github.com/thedatalife/sketch-ball

Which part do you keep for yourself?