Let the model write the system

The short version

• I had 500 messy files to clean up.
• Asking the AI to fix each one cost about 131,000 tokens.
• Asking it to write a program that fixes them cost about 5,500.
• The program got more of them right.

The job

Turn this:

{ timeout: "90s", endpoint: "api.internal:8080", enabled: "yes",
  owner_note: "checkout breaks if this is down" }

Into this:

{ timeoutMs: 90000, host: "api.internal", port: 8080, enabled: true,
  priority: "urgent" }

Most of it is simple rules. "90s" is 90,000 ms. "yes" is true. A program can do that.

One part is not a rule. Priority comes from a note a person wrote. "Checkout breaks if this is down" means urgent. That needs a model.

What I tried

Here is the idea of the program it wrote, cut down:

const configs = tool("configs", {});
const notes = configs.map((c) => c.owner_note);
const priorities = tool("judge_priorities", { notes });

configs.map((c, i) => ({
  timeoutMs: parseTimeout(c.timeout),
  ...parseEndpoint(c.endpoint),
  enabled: parseEnabled(c.enabled),
  priority: priorities[i],
}));

The lesson

Writing a program alone did not save much. Number 2 proves that.

What worked was splitting the job:

• Rules go into code, written once.
• Judgment goes to a small, cheap model, all in one batch.

Then the big model's cost stays the same whether you have 500 items or 50,000.