Structured extraction into a table you can query
Lesson 2 of 6 · Literature at scale
1 · Learn the move · Template pattern
Reading twenty papers and remembering them as vibes is how citation errors are born. The alternative is structured extraction: a fixed schema applied identically to every paper, producing a table you can sort, filter, and query later. The template does the discipline: fields for system, method, sample size, effect direction, conditions, and - the load-bearing one - a supporting quote with its location for every substantive claim. Extraction rules are strict: each field filled ONLY from the text you supplied, [NEEDS: ...] where that text does not state a value, and no reaching across papers to fill one paper's row. Every row carries its source and a verification state - VERIFIED only once you confirm the quote sits where the table says. The table is your queryable memory; the quotes make it citable rather than merely plausible.
Extract this paper into my evidence table using the fixed schema: SOURCE (authors, year, DOI as given) / SYSTEM / METHOD / N / EFFECT (direction + magnitude as stated) / CONDITIONS / KEY QUOTE (verbatim, with section) / LIMITS (as stated by the authors) / STATE (starts UNVERIFIED; I flip it after checking the quote). Rules: every field from the supplied text only; [NEEDS: ...] where the text is silent; the quote must be verbatim and located, not paraphrased; if the supplied text is abstract-only, the row says ABSTRACT-ONLY in METHOD. One paper, one row, no blending. [paste text]
2 · Your turn. You write the prompt
Fourteen papers on your extraction buffer's compatibility with downstream assays are read and annotated, and the knowledge currently lives in marginalia across fourteen PDFs. Build the table that makes them one queryable object - with quotes located, so December-you can cite without re-reading.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.