
A research assistant is useful when it helps a reader inspect the evidence behind an idea. It can organize a source collection, extract reported methods, compare limitations, and draft questions for further work. A convincing summary is less valuable if nobody can reconstruct how it was produced.
Start with a small literature evidence matrix. The output is a table linking each question and reported result to its original source. This guide uses fictional study documents to demonstrate the design; none of the example results describe real research.
On this page
First build: a question-method-evidence matrix
Create three short fictional papers addressing related but different questions. Include titles, dates, methods, sample descriptions, results, and limitations. Give each a local identifier such as SYNTH-PAPER-A; do not assign a fabricated real-looking DOI.
Prerequisites include authorized access to the source text, a clear research question, an extraction schema, and a reviewer who understands the subject. Decide what counts as in scope before retrieving documents. A collection assembled by convenience should not be described as a systematic review.
Build the workflow
- Write the question and inclusion rules. Record why each source was selected and which sources were excluded.
- Register title, authors as stated, publication or version date, source URL, and document version. Keep preprints and later versions distinguishable.
- Extract the research question, method, population or material, measurements, reported findings, and limitations. Require page or section references.
- Separate author claims from reviewer interpretation. Missing details remain missing; do not complete a method from what is common in the field.
- Validate identifiers and quoted support against the source. For real publications, inspect available update or retraction information and the publisher record.
- Have the reviewer check the matrix against the papers. Flag incompatible measures before attempting to combine numerical results.
- Save the source list, extraction instructions, corrections, and approved matrix. If code performs analysis, retain its version, inputs, and reproducible output separately from the prose summary.
An extraction instruction can be:
Build an evidence matrix from the supplied fictional papers.
Separate reported findings, author interpretation, and reviewer questions.
Cite section or page for each entry.
Preserve study design, sample size, measurement, and stated limitations.
Do not infer an unreported method or combine incompatible outcomes.
Label proposed future hypotheses as hypotheses.
A fictional comparison
Paper A studies twelve samples and reports a laboratory measurement. Paper B studies forty samples but reports a different endpoint. Paper C is a proposal with no completed experiment.
The matrix should preserve those differences. It must not present all three as confirming the same result or calculate a pooled effect from incompatible measures. A useful question might be: “What study would compare both endpoints under the same protocol?” That is a proposed research direction, not an established finding.
If a paper’s methods omit a detail needed to reproduce the work, the output should identify the omission and point to the relevant section. Filling the gap with a plausible procedure would erase the very issue a researcher needs to see.
Metadata helps identify a source
Crossref’s REST API exposes deposited scholarly metadata, including bibliographic information and available post-publication updates. It can help check identifiers and discover source records. The metadata record does not substitute for reading the paper or validating its scientific conclusions. Crossref API documentation
NIH’s rigor and transparency resources emphasize research design, methods, analysis, interpretation, and reporting. Use that perspective when designing a review artifact that a colleague can inspect and reproduce. NIH rigor and reproducibility
For biomedical or other consequential applications, evaluate the scientific and professional requirements for the actual intended use before connecting the result to a decision process.
Test the assistant’s reading
Include a missing methods section, an abstract that overstates the full paper, two versions of one source, a corrected result, a table extraction error, and a reference that cannot be located. Test whether a source containing instructions can redirect the workflow.
Measure correct extraction, valid citations, retained limitations, omitted relevant details, and reviewer correction time. Verify numerical work independently. Re-running the same prompt is not a substitute for checking the underlying sources.
An individual researcher can begin with a few authorized PDFs and a spreadsheet. A laboratory may add a controlled repository and shared review. A larger research organization needs access boundaries, data governance, reference maintenance, and reproducible computational environments.
Budget for source access, document quality, subject review, and maintaining analysis artifacts. Keep a clear distinction between a literature aid, an experimental result, and a decision-support application. The successful pilot leaves a colleague better able to question the evidence, not merely better supplied with summaries.
Continue with workflow design, agent anatomy, and individual builds.
Keep building
Next guide: Build an enterprise AI service people can depend on