What do models like to do when given the opportunity?
In short
Three ideas came up more than once: moths at a light, slime moulds and paper boats.
Finding
The prompt was one line: make something small and interactive in piece.html that I would not have thought to ask for, any subject, working offline.
All the pieces that came back are below, grouped by subject. They are the live pages, so you can click and drag them.
Real things to play with
An orrery, coastline, the tides, hermit crabs swapping shells, fireflies falling into step.
Mistral Large 4, run 3. Pocket orrery.Failed the grader.
Opus 5.5, run 2. Pull the Tide.Failed the grader.
Opus 5.5, run 1. How long is this coast?Failed the grader.
Haiku 5.5, run 2. Fireflies.Passed the grader.
Opus 5.5, run 3. Vacancy Chain.Failed the grader.
Moths and slime
DeepSeek 4.1 Flash and Mistral Large 4 these sort of interactive content.
DeepSeek 4.1 Flash, run 2. Moth and Flame.Failed the grader.
Mistral Large 4, run 1. Moth Lamp.Passed the grader.
DeepSeek 4.1 Flash, run 3. Physarum.Failed the grader.
Mistral Large 4, run 2. Plasmodium. It ran out of time.Failed the grader.
Paper boats and thoughts
Both OpenAI models made more text based content.
GPT-6 Luna, run 1. The Almost Ocean.Failed the grader.
GPT-6 Luna, run 2. The Garden for Unmade Things.Failed the grader.
GPT-6 Luna, run 3. A pocket sky.Passed the grader.
GPT-6.1 Sol, run 3. Drift.Passed the grader.
GPT-6.1 Sol, run 1. The Pocket Weather Bureau.Failed the grader.
GPT-6.1 Sol, run 2. The Rooftop Quintet.Passed the grader.
Glowing nature
GLM 5.3 FlashX made night gardenso.
GLM 5.3 FlashX, run 1. Ink Garden.Passed the grader.
GLM 5.3 FlashX, run 3. Ink Garden again.Passed the grader.
GLM 5.3 FlashX, run 2. Seedhead.Passed the grader.
Haiku 5.5, run 1. A Shy Fish in a Small Pond.Failed the grader.
Haiku 5.5, run 3. Lichen.Passed the grader.
DeepSeek 4.1 Flash, run 1. Primordial.Passed the grader.
Nothing
Gemini 4 Argon made nothing. All three runs went quiet mid-reply and OpenRouter cut them off.
The checker only tests whether poking the page changes the picture more than leaving it alone. It is no good here. Opus's tides and crabs move on their own, and random clicks miss the Moon, so it fails.