A research agent should reduce clerical work while keeping the evidence visible. A useful first assignment is a small literature brief: identify relevant papers, record what each actually reports and make it easy to return to the original passage.
Do not start by asking it to settle a field’s most disputed question. Start with a narrow question and an output you could audit in an hour. For example: “Which of these five supplied papers measures this outcome, in which population, using which method?” That scope gives you a way to distinguish a missed paper from a misleading interpretation.
Define the question and the exclusions
Write down the topic, date range, study type and exclusion criteria before searching. These are decisions about the brief, not something the agent should silently invent. If you change a criterion after seeing results, record the change so another person can understand the resulting selection.
For an exploratory reading exercise, give the agent a small set of public papers you already selected. A systematic review requires a different, documented method, including appropriate search coverage and screening. An attractive automated summary does not turn an informal search into a systematic review.
Extract before synthesizing
Ask for a table with one row per paper: title, authors, publication year, stable identifier, research question, method, sample, reported result and limitation. Include a source location for each substantive field. Require “not reported” where the paper gives no answer.
Open every cited paper. Check the identifier and publication details first, then compare the table with the methods and results. An abstract may leave out a limitation that matters to your question. A source link pointing to a real paper is still insufficient if the reported number came from somewhere else.
Only after checking extraction should you ask for a synthesis. Request agreements, disagreements and reasons the studies may not be comparable. Keep an observation such as “these papers use different outcome measures” separate from a claim that one intervention works better. The latter needs evidence that supports that comparison.
A useful first prompt
“Using only these supplied public papers, make an extraction table for this question. Cite the section or page supporting each result. Distinguish authors’ findings from your interpretation. Flag missing information and incompatible measures. Do not add references, run analyses or access unpublished data.”
Inspect one row thoroughly before commissioning the whole table. This catches an unsuitable extraction format early. Save your corrections beside the table so you can tell whether a revised version fixes the underlying mistake or merely changes its wording.
Protect the work that is not public
An initial literature task does not require raw participant data, peer-review manuscripts, grant applications or laboratory credentials. Before providing any of those, follow your institution’s access, confidentiality and data-handling requirements. De-identification is a substantive research decision, not a label to apply casually to a file.
If you later use an agent to write analysis code, review and run that code in the appropriate controlled environment. Check inputs, transformations and outputs. Keep a record of which results were actually calculated; a plausible numerical answer in a chat is not a performed analysis.
Authorship and the final record
For journals following ICMJE recommendations, its guidance on AI use by authors places responsibility on human authors and does not treat an AI tool as an author. Check the target journal’s disclosure requirements. Do not assume a single policy covers every publication.
Keep the question, selected sources, extraction table, corrections and final synthesis together. A useful agent-assisted brief lets a colleague inspect the evidence without repeating the entire conversation. Measure success by accurate extraction and visible uncertainty, not by how many papers the tool claims to have read.
