Ground everything
The single most important rule for research work: supply the sources. A model answering from memory produces confident, unsourced, partially invented text. The same model given three documents and told to answer only from them, citing the section for each claim, is dramatically more reliable — and its failures become visible, because a missing citation is easy to spot.
Add the instruction explicitly: "Answer only from the supplied text. If the answer is not present, say 'not stated in the sources'." Models comply with this far more readily than people expect.
Summarising for decisions
A summary that shortens without prioritising is not useful. Ask for the shape you actually need:
- Decisions made, with owner and date
- Open questions and who must answer them
- Claims made, each with the evidence offered
- What changed since the previous version
AI Summarizer and Meeting Notes apply this framing by default rather than producing a shorter version of the same prose.
Structured analysis
For comparisons and evaluations, define the criteria before you see any output, and require a table. This blocks the model's habit of praising whatever was mentioned last. Then require an explicit recommendation with a stated assumption and a stated risk — hedged conclusions are the most common failure of AI analysis.
Stress-testing conclusions
Once you have a conclusion, attack it in a fresh thread: "Here is a conclusion and its supporting reasoning. Argue the strongest case against it. What evidence would change the answer?" A fresh thread matters, because a model that just wrote the argument is anchored to defending it.
Traceability
Keep the sources, the prompt and the output together. Six weeks later the only question that matters about an analysis is where a number came from — and "the AI said so" is not an answer you want to give. Research Assistant keeps runs in your private history for exactly this reason.