AI Research Workflow: Write a Brief with Verifiable Sources

A useful research brief answers a bounded question and shows where the answer comes from. It should make disagreement easier to investigate. A long synthesis with an impressive bibliography is less useful if its central claim cannot be traced to a particular passage.

Define the decision and the stop condition

Write one question, a reader and a deadline. For example: “Should our fictional support team pilot an evening help channel next month?” Define what evidence would change the decision: request timing, available staffing and the cost of a limited trial. Exclude unrelated market trends unless they affect that decision.

Gather sources into a ledger before drafting. Start with the relevant original records, official documentation or underlying study. Web extraction tools such as Firecrawl can assist collection, but they do not establish reliability. Record source date and scope so that an old regional result does not become a claim about today's whole market.

Use a claim ledger

Claim | supporting passage | source URL or record ID
Date and population | limitation | conflicting evidence
Status: supported / qualified / unresolved
Implication for the decision

Fictional evidence: 30 of 200 tickets arrived after 18:00, but 18 of those were automated notifications. “15% of customers need evening help” is unsupported: tickets are not customers and notifications are not requests for assistance. The narrower supported observation is 12 potentially relevant tickets, pending classification.

Draft from accepted claims only

Ask for a one-page brief with a recommendation, supporting evidence, the strongest counterargument and the next information to collect. Require every factual sentence to point to a ledger row. Keep assumptions visibly separate. If the ledger cannot answer the question, the recommendation can be a small experiment rather than a definitive purchase.

A good conclusion for the example is to classify the twelve tickets and estimate staffing before choosing an evening channel. It is not to announce proven demand. This distinction prevents fluent writing from converting incomplete evidence into certainty.

Review for traceability and decision value

Choose three claims at random and follow each back to its source. Check denominators, units, dates and whether the quotation supports the actual wording. Ask a colleague to argue the opposite conclusion using the same ledger. Revise if the brief ignores a material limitation.

Publish the brief with its source ledger and review date. Revisit it when the decision changes, not merely because more documents become available. For repeated external collection, compare the Firecrawl assessment; for research priorities and reading habits, use the researcher guide.