Like the rest of the world, I am very impressed by the long-term capabilities of Anthropic’s latest model. Having a little usage to spare, I decided I would try it on a cute little idea that I have been sleeping on.
The idea: a fixed camera and a projector, both pointed down at a paper jig on my desk. Drop a handwritten exam answer into the jig, a local VLM grades it, and the projector draws the marks — ticks, the score, a one-line remark — back onto the paper, registered to the right spot on the page. I wrote a spec, handed it to Fable, and asked it to produce a working MVP. The whole thing came together in a single conversation.
What Fable can do
On its own initiative, Fable did some impressive things that I wouldn’t have bet an AI would be able to do:
- Calibrated the jig by superimposing a grid and reading off the coordinates.
- Calibrating the projector by projecting dots and differencing frames to find them.
- Generating its own synthetic test data with realistic transforms.
- Debugging a MacOS window-handling issue.
- Figuring out color-channel issues
It worked autonomously for about forty minutes, and had a working prototype. It built the product up by layer by layer, instead of all at once, and built up a set of experiments and tests to validate at each step of the way.
My hands
My contribution to the endeavour was the design and assembly of the jig, some mild discussion in writing the spec, and swapping in pages on request.
The full source is on GitHub, and Fable’s own unedited writeup on the session is fascinating. It almost (but not quite) thinks like a human would.