Skip to main content

Research & Synthesis

All levels
What you'll learn
  • Use the five research moves that actually pay off — summarize, compare, map, extract, pressure-test
  • Write a grounded prompt that forbids outside knowledge and demands a supporting quote per claim
  • Spot a fabricated citation before it reaches your document
  • Run the five-step research flow, from gathering sources to spot-checking the critical facts
  • Keep stated facts and model inference visibly separate in your notes

AI is a superb research accelerator — and a dangerous one if you trust it blindly. Used right, it compresses hours of reading into minutes; used carelessly, it confidently invents sources. Here's the disciplined version.

High-value research moves​

  • Summarize a long document/paper into key points + implications.
  • Compare multiple sources into a table (what they agree/disagree on).
  • Build a landscape — map the players, approaches, or arguments on a topic.
  • Extract specific facts/figures from a document (with citations).
  • Pressure-test — "what's the strongest counterargument to this?"

The grounding discipline (non-negotiable)​

Grounded summary — quote-per-claim, 'not stated' allowed

Using ONLY the text I provide, summarize the key findings and quote the exact
sentence supporting each one. If something isn't in the text, say "not stated."
Text: """[paste]"""
  • Feed it the sources. Don't ask it to recall facts from memory for anything important — paste the material or use a tool that retrieves it (RAG).
  • Demand citations. "Quote the sentence that supports each claim." Unsupported claims become visible.
  • Separate fact from inference. Ask it to label what's stated vs. interpreted.

The third one is the move most people skip, and it is the cheapest. A summary that mixes "the paper reports X" with "which suggests Y" reads fine and is useless for checking, because you can no longer tell which half you need to verify.

Fact vs. inference — force the two apart

From the text below, produce two separate lists:
STATED — claims the text makes explicitly, each with the exact supporting sentence.
INFERRED — anything you concluded that the text does not say outright, each with
what it is based on and how confident you are.
Put nothing in STATED that you cannot quote.
Text: """[paste]"""
Fabricated citations are the classic trap

Models invent realistic-looking papers, authors, and URLs. Verify every citation exists before you rely on it. See Hallucinations.

A fabricated citation rarely looks wrong. It looks exactly right: a plausible author pair, a real-sounding journal, a year that fits. Which is why the only reliable check is mechanical — resolve the citation, don't evaluate it. Three signals that something needs resolving before you cite it: a DOI or URL that 404s, a title you cannot find anywhere but the model's answer, and a paper quoted for a claim that is slightly too perfect a fit for your argument.

A research flow​

Guided walkthrough1 of 5
  1. Collect the material, or point a retrieval tool at it. This is the step that decides whether the rest of the work is grounded or improvised.
Pro tip
  • Ask for the strongest counterargument in a separate turn, not as part of the summary. Bundled together, the model tends to produce a counterargument it has already defeated.
  • Keep the source text in the conversation for as long as you are asking questions about it. Once it scrolls out of context, answers quietly shift from reading to remembering.
  • When comparing sources, give each one a short label (S1, S2, …) and require every claim to carry its label. Unlabeled claims are the model's own voice, and they are now easy to spot.

Check yourself​

Check yourself

0/3
  1. What does the 'not stated' instruction in the grounding prompt actually buy you?
  2. A model gives you a citation with a convincing title, two plausible authors, and a journal you recognize. What is the correct next step?
  3. Why summarize sources one at a time instead of pasting all five into one prompt?

Takeaways​

Key takeaways
  • Feed the sources; never let an important fact come from the model's memory
  • Demand a supporting quote per claim, and sanction 'not stated' so gaps become visible
  • Keep STATED and INFERRED in separate lists — mixed, neither can be checked
  • Resolve citations mechanically. A fabricated one looks exactly right by design
  • Report disagreement between sources instead of asking the model to settle it
  • Spot-check the facts your conclusion depends on, in the original

Next​