What Should You Give an AI Research Agent Before It Starts Searching?

Research agents often fail before the first search. The user gives them a topic—“AI and student learning”—and expects the agent to infer the population, evidence standard, date range, exclusions, and final output. The result is long because the task was vague, not because the search was good.

A useful agent brief should make seven decisions visible.

Write a question, then say what you will do with the answer. “How does AI affect learning?” is broad. “Should our writing center pilot AI feedback for first-year students?” gives the research a decision.

Name who, where, and what counts. Specify whether “students” means undergraduates, medical trainees, or K–12 learners. Define whether “AI feedback” includes grammar checkers, chatbots, automated scoring, or only generative feedback.

Give the agent one to three papers that are clearly relevant, plus known authors, acronyms, and alternative terms. Seed papers show what “relevant” looks like and provide references to follow. Tell the agent not to treat them as automatically correct or complete.

State dates, languages, publication types, fields, and exclusions. If preprints are allowed, ask the agent to label them. If you need empirical studies, exclude commentaries and product pages.

Ask for an artifact you can inspect: a paper table, evidence map, annotated reading list, or claim-to-source report. Define the fields. “Write a literature review” hides too many decisions inside prose.

Require stable links, DOI or publication records, source locations for important claims, and a search for studies that disagree. Tell the agent to write “not found” or “not reported” rather than filling gaps.

Say when the agent should stop: after a time limit, after a set of diverse relevant papers, or when new searches stop changing the evidence map. Name decisions it cannot make, such as final inclusion, ethical approval, or running code.

This brief will not guarantee a perfect search. It makes failures easier to see and correct.

You can run the question in WisPaper Search or give a longer version to Scholar Agent. Save the approved candidates to My Library, then use Library QA for questions based on that stable set.

The brief should travel with the project. When the search changes direction, update the brief rather than relying on the agent to remember an implicit decision from twenty messages ago.