How I Screen 100 Papers Before Lunch with Elicit
A repeatable Elicit workflow for literature screening: extract columns that matter, trust the data table over the chat, and know exactly when to stop believing it.
Literature screening used to be my least favorite kind of homework. Elicit turned it into a sorting problem, and this is the workflow that gets me through an initial screen of a hundred papers in a morning.
Step one: start from a seed paper, not a keyword. I paste in the one paper my advisor actually trusts and hit 'Find similar work'. Keyword search in academic databases punishes you for not guessing the exact jargon a field uses; seed-paper search does not.
Step two: build the data table before reading anything. Elicit's extraction columns are the whole product. For my last review I set columns for sample size, intervention type, reported effect size, and study duration. Then I add the column that saves the most time: 'limitations stated by authors'. Sorting that column surfaces the weak studies instantly.
Step three: read the table, not the papers — at this stage. I sort and filter until roughly thirty papers survive, and only then do I read those in full. The extraction is not perfect: maybe one in fifteen cells is wrong or missing, which is why every number that survives into my notes gets checked against the actual PDF, one click away.
The honest caveat: Elicit works best for empirical studies with clear methods. My colleague in literary theory got little value from it — there is no sample size to extract in a close reading of a novel. For STEM screening, it is the fastest tool I have used.
Step four, and this is the step people skip: export the table to CSV before you close the tab. Elicit notebooks are fine, but my review lives in a spreadsheet I control. Tools change; your data should not be hostage to them.
Frequently asked questions:
Q: Is Elicit's extraction reliable enough to cite directly?
A: No. It is a screening tool, not a citable source. Every figure that appears in my own writing gets verified against the original paper. The table decides what gets read, never what gets cited.
Q: How does it handle paywalled papers?
A: It extracts from abstracts when the full text is paywalled and marks the extraction as lower confidence. For a proper screen of paywalled studies, I route through my university library access first.
Q: Can it do the whole review for me?
A: It cannot, and you would not want it to. It removes the mechanical part — finding, sorting, tabulating — which is maybe 60% of the hours. The judgment part is still yours.