The prompt was the wrong unit of work

We just shipped something that contradicts the conventional wisdom on prompting Claude.

The conventional wisdom is that better output comes from better prompts. Learn the patterns, tune the wording, iterate. It works. It also doesn't scale past one person at a keyboard.

An Anthropic engineer put the alternative bluntly in a video going around this week: “You’re not supposed to prompt Claude. You’re supposed to build a system that prompts itself.”

That matched what we found the hard way.

At Foltrigg, our evidence discovery product running in the Shelby County DA’s office, the first version put a very good search box in front of prosecutors. It worked, and almost nobody used it well. Prosecutors don’t want to phrase a query. They want the gold nuggets — our primary DA contact used that exact phrase. The prompt was the wrong unit of work.

So we stopped shipping a place to type and started shipping the pipeline. Evidence lands, gets classified by type, and the system asks its own questions of it — every time, whether or not a human thinks to. One evidence processor there watches body cam footage second by second to cut the relevant segments. That is not a prompting problem. That is work the system should have finished before she opens the file.

Same pattern at Cosmos OS. Our data room runs 73+ specialized agents across M&A diligence. Nobody prompts 73 agents. They fire when documents arrive, and the analyst reads the output.

Three things I’d tell anyone building this way.

The context you load before the first token matters more than the prompt itself. Get the file structure, the schema, and the definitions right and the model stops guessing.

Triggers beat interfaces. If the work only happens when someone remembers to ask for it, it mostly won’t happen.

Ship the boring part. Retries, evals, logging, permissions, audit trail. That’s where an assistant turns into infrastructure.

The caveat: this only holds when the task is repeatable and the output is checkable. Novel judgment still wants a person at the keyboard, and in a courtroom that person is legally required to be there. We’re not automating the decision. We’re automating everything upstream of it.

If a workflow only produces value when someone remembers to prompt for it, you don’t have a product yet. You have a demo.

Pick one recurring task this week and ask what would have to be true for it to run without anyone opening a chat window.

What’s the workflow at your company that should already be running itself — and what’s actually stopping it?