Claude Became the Data

Running a few unassisted Claude Code instances against a hackathon task — same task, different setups, see what shakes out.

One of them almost gets it. Not perfect, but closer than the others by a meaningful margin. The difference isn't model, isn't context window, isn't some clever system prompt I wrote.

Before I turned it loose, I had it manually do the exercise. As if it were an employee given the job. No scaffolding, no instructions about what the code should do — just: here are the JSON docs, now do the thing a person would do.

Claude became the data.

And then, when I actually set it to work on the task, it had already inhabited the problem. It had the shape of the work in its context — not described, not explained, but done. There's a difference between a model that's been told what a thing is and a model that has just spent several exchanges being that thing.

This is probably obvious in retrospect. Every prompt engineer has some version of "show don't tell" as a rule. But there's something weirder happening here — the grounding isn't an example you fed it, it's the model's own output, its own reasoning, its own approximation of the work.

It briefed itself. And then it believed it.