LLMs as compilers
Outsourcing thinking and amplifying thinking are not the same thing. The difference shows in the prompts.
March 2026
The thinking happens before you open the prompt.
A TypeORM rewrite: branch strategy, module mapping, 11 execution steps, 25 files to create, 7-step verification checklist — all authored before Claude saw it. The prompt was the spec. Claude was the compiler. The cognitive work was in the decision, the architecture, the structure. The typing was downstream.
A compiler doesn't write your program. It realizes it. The spec is the thinking. The output is the trace.
The prompts are the evidence. An auth edge case in a government portal: hard to reproduce, specific behavioral analysis, mental model of the state transitions built from hours of navigating the actual system. No LLM can invent that. The domain knowledge is irreducibly yours. The prompt is the analysis. Claude writes the handler.
A POC idea that emerged mid-session — recognizing that a workflow a client already ran manually was structurally identical to something the code was already doing. That's the insight. The connection between two domains, the recognition of structural similarity, the question worth asking — that's the work. The LLM builds what was conceived. The conception is not delegatable.
Outsourcing thinking and amplifying thinking are not the same thing. One produces output that could have come from anywhere. The other produces output that could only have come from you — because the context, the decision-making, the domain knowledge are irreducibly yours. The LLM executes. You architect.
The difference shows in the prompts. Vague input, vague output. Specific input — branch strategy, module table, verification sequence, the edge case you spent three hours reproducing — specific output. The prompt is the program. The rest is compilation.
"Did you use AI?" is the new "did you write it by hand?" Both questions locate value in the wrong place. When you can't evaluate the output, you audit the process. That's not a principled position about quality. It's a confession that the evaluator doesn't know what quality looks like in this domain.
The prompts are the work. Read them.