Give a research agent an evidence contract and a finish line
Define the decision, approved sources, weak-evidence flags, output format, and stop condition before a long-running agent begins.
What matters today
Copy a long-horizon research prompt that controls sources, flags weak evidence, and delivers a one-page decision brief with a clear finish line.
Key points
- Long research runs need a specific decision, approved sources, evidence rules, and an observable finish line.
- Every material claim should cite a source and location; weak, single-source, conflicting, and missing evidence should remain visible.
- A one-page decision brief should state the recommendation, confidence, open questions, reversal condition, and next owner.
- A claim ledger preserves traceability without turning the decision brief into a research transcript.
- Research permissions should remain read-only until a separate approval authorizes an external action.
The executive summary remains public for search engines and AI answer systems. Subscriber-only sections are marked separately on the page.
Article roadmap
What you will learn
-
A research prompt for long runs in Grok Bot, Claude Cowork, or ChatGPT Work
-
How an evidence contract controls sources, conflicts, freshness, and confidence
-
How to turn a broad goal into the minimum set of decision questions
-
What a one page brief must contain before the agent can stop
-
How to test the result without rereading the entire research trail
Jordan, a strategy director, asks an agent to compare three payroll vendors before Friday. Six hours later, the agent has opened dozens of pages and produced a polished market overview. The brief never answers the decision that matters: which vendor can support the company's countries, security requirements, and launch date within budget.
All that activity hides the failure. The assignment gave the agent a subject, but no evidence boundary or finish line. It spent six hours circling the decision instead of answering it.
A useful assignment reads like a research contract. It names the decision, limits the evidence, keeps uncertainty visible, and describes the artifact that must pass review. That contract works in Grok Bot, Claude Cowork, or ChatGPT Work even though each product handles permissions differently.
Pro subscriber access
This briefing is reserved for Pro subscribers.
Enter the email tied to your Beehiiv Pro subscription to continue.
Three deep dives. Four useful moves. One email worth opening.
PromptHacker turns the AI firehose into practical next steps for work, health, family, and everything time keeps trying to steal.