A practical workflow for researching with your own sources
Four steps. 1) Attach the source and ask for claim, method, sample and stated limitations. 2) Ask what it does not establish — the gap between the result and the headline. 3) Add a second source and ask exactly where the two diverge and on what evidence. 4) Ask for the page or section behind each claim, then open the document and check it. Never ask a model what the literature says; ask it about text you supplied.
The dangerous way to use AI in research is to ask it what the literature says. It will answer fluently from memory, and some of that answer will be invented.
The reliable and genuinely fast way is to confine it to documents you attached. What follows is that workflow, step by step, with the prompts and the verification checks each step needs.
Step 1 — Question one source properly
Attach the document and ask for claim, method, sample and the limitations the authors state themselves. Ask for the supporting sentence alongside each item; that single instruction is what makes the output checkable rather than plausible.
Step 2 — Ask what it does not establish
The interesting gap is usually between what a study measured and what its abstract implies. Ask what the paper assumes without arguing for it, and what its result would not license you to claim.
Step 3 — Compare sources against each other
Bring a second source into scope and ask where exactly the two diverge and on what evidence. For a larger corpus, put the documents in a knowledge base first so passages are retrieved from the whole set rather than one attachment at a time. Branch the thread when a tangent deserves following without derailing the main question.
Step 4 — Verify before you rely on anything
- Ask for the page or section behind each claim, then open the document and read it.
- Treat anything it cannot locate in an attachment as unsupported.
- Re-check every number and quotation; transcription and table reading are where errors hide.
- Read anything load-bearing in the original — a summary flattens nuance by design.
Don't regenerate what you already asked
Search your own message history by keyword to recover the summary you produced three weeks ago, rather than regenerating it and getting a subtly different answer.
Where it fits
Triage a reading pile
Claim and method per paper, to decide what deserves a full read.
Compare sources
Two papers in scope, one question: where do they conflict?
Stress-test a thesis
Ask for the strongest objection to the argument you are building.
Example prompts
- From the attached paper only: claim, method, sample size, and the limitations the authors state.
- These two papers reach different conclusions. Where exactly do they diverge, and on what evidence?
- Argue against my summary. What would a hostile reviewer say?
Limitations
- No web browsing or live search at any step — you supply every source.
- It can produce plausible citations that do not exist; never accept a reference it was not given.
- Summaries flatten nuance, so anything load-bearing should be read in the original.
Frequently asked questions
What is a reliable step-by-step way to research with AI?
Question one source at a time for claim, method and limitations; ask what it fails to establish; compare it against a second source; then verify each claim against the page it came from. Confining the model to attached text is what makes the process trustworthy.
How do I stop AI from inventing citations?
Never ask what the literature says. Attach the sources, instruct it to answer only from them, and require the supporting sentence for each claim so anything unsupported becomes visible immediately.
How should I handle a large set of documents?
Attachments are capped at 5 files per message, 15 MB each, so a large corpus belongs in a knowledge base: upload once, and relevant passages are retrieved into any later conversation.
What must I still do myself?
Judge the quality of the evidence, check every number and quotation in the source, and read anything your conclusion rests on in full. The model compresses text; it does not evaluate it.
Run this on the paper on your desk
Attach it, ask for claim and limitations, then check the page references.
Try this workflow