One prompt clause turned my analysis agent into an experimentalist.

I've run a lot of analysis sessions with Claude over the past year. Financial models for Mile High Golf, hardware tradeoffs for Quantum Caddy, literature sweeps for Parley. The output quality always lands somewhere between useful and unusable, and for months I treated that variance as luck.

Then the July 5th physics session happened, the one I wrote about in the simulator teardown. That session came back with runnable Python models, a dozen figures, CSVs of every fit, and a ranked list of six experiments that would retire the remaining uncertainty. Not a single unquantified claim in three hours of work. It was the best analysis output I've gotten from any model, and none of it was luck. The difference was the brief.

The clause that does the work

The brief was one page. Most of it was domain setup. But one paragraph changed the character of everything that came back:

“You work like an experimentalist, not a theorist with a whiteboard: no physical claim survives without a number, an error bar, and a plot. You write code that another engineer can run and reproduce. When the data can't support a claim, you say so and state what measurement would settle it.”

Read that again as a contract, because that's what it is. It sets an evidence standard (number, error bar, plot), a reproducibility standard (code someone else can run), and an honesty standard (name the missing measurement instead of guessing). Every deliverable from the session can be audited against those three clauses. That's the whole trick.

Anatomy of the one-pager

The full brief had six parts. None of them is exotic. The discipline is in writing all six instead of two.

- A role with an epistemic posture. Not "you are a physics expert" but "you are a quantitative sports-physics research engineer" plus the experimentalist contract above. The role says what the agent knows; the contract says what counts as done. - The domain described physically, not as a data dictionary. I wrote two paragraphs on the actual regimes of the problem, what happens in the air versus what happens at the surface, and where the interesting behavior lives. I did not paste a schema and hope. - The data grant, with its limits stated up front. What files exist and what they contain, but also what they are not. I told the session explicitly that no measured throws existed. - A phase plan. Reproduce what we built, then correct it, then compare, then budget the uncertainty. Order matters and I chose it, not the model. - A deliverables contract. Scripts, figures, CSVs, and a final written audit. Named formats, so "analysis" couldn't mean a wall of prose. - The honesty clause. Permission and obligation to say "the data can't support this," paired with the requirement to name the experiment that would.

What each clause bought

The role and contract bought me zero speculation. Everywhere the data ran out, the session said so in plain language and moved the question to the measurement list instead of papering over it.

The physical domain description is the part I'd defend hardest. Because the brief framed the problem as two coupled regimes instead of one dataset, the session went looking for trouble at the boundary between them. That boundary is exactly where it found the defect that mattered. A schema-only brief would never have pointed there.

Stating the data's limits up front inverted the usual failure mode. Instead of trusting the data too much, the session opened by writing a table of what each file could and could not support, and flagged that every tunable in our grid was an assumption rather than a measurement. That table set the honesty budget for the whole session.

The deliverables contract is what made the output survive the night. I can re-run every fit. Curtis can re-run every fit. When we measure real throws next month, the scripts take the new data without ceremony.

And the honesty clause produced the most valuable artifact of all: a prioritized list of experiments, each tied to the amount of prediction error it retires. The session ended by telling me what it couldn't know and exactly what it would cost to know it. I have hired consultants who never once did that.

Questions get answers. Contracts get work.

The pattern underneath is that I stopped writing prompts shaped like questions and started writing prompts shaped like engagements. A question invites an answer, and answers from a capable model arrive fluent and confident whether or not the evidence is there. A contract invites work product, and work product has to show its receipts.

The reusable skeleton, for any deep-analysis session:

- Role plus epistemic posture (what counts as a claim). - Domain described the way a practitioner would describe it to a new hire. - Data grant with explicit limits. - Phase plan in the order you want the reasoning to happen. - Deliverables by name and format. - Honesty clause with a required missing-measurement list.

The brief took maybe forty minutes to write, most of it on the domain paragraphs. The simulator it audited took a weekend and a GPU. I know which one I'd pay for again.