Sign in Subscribe
3 min read

Build a Claims Bank Before You Write Anything

The highest-leverage hour of an AI-assisted writing project isn't spent writing — it's spent turning your literature into one file: one claim per line, every reference traceable. The cure for hallucinated citations.
phd-compass header graphic reading "Build a claims bank" — AI literature review workflow, showing Extract → Dedupe → Verify.
Give the AI a citation supply chain, not a citation memory.

This comes from a side project I've been toying with: several PhD students are attached to me, and I keep asking myself how to make their writing lives less painful. The single highest-leverage hour of that project was not spent writing. It was spent turning a pile of literature into a claims bank: one file where every fact from the reading lives in the same format, attached to its full reference.

Every draft we produced afterwards cited from that one file. Every citation could be traced back mechanically. When you hear that AI invents references, this is the cure.

The format

One claim per line, always the same shape:

The claim, stated fully with its numbers. (Author et al., Year) [Full reference: authors, year, title, journal, DOI]

The full reference travels with the claim. You will never again ask "where did I read that?" — the same discipline behind the Reference field in the 6-field literature note system, scaled up for AI-assisted work.

📄 Free kit: The claims bank feeds writing the same way good notes do. "From Notes to First Draft" is my free kit for exactly that step — the note-to-paragraph workflow and worksheets I use myself. Grab it here (free, members only).

The steps

  1. Gather your source reviews or papers. I started with a handful of literature reviews in the project's field. PDFs are fine; have the AI extract the text first.
  2. Extract exhaustively, in parallel. Prompt per document: "Extract EVERY substantive empirical claim from this review. One bullet per claim, exact numbers preserved, resolve every citation number against the reference list, output in this format: ..." Expect 100+ claims per review.
  3. Deduplicate in stages. Rule: two bullets are duplicates only if they assert the same fact from the same source study. Keep the version with the most precise numbers. If the same fact comes from different studies, merge into one bullet citing all sources. Merging the reviews cut several hundred raw claims down by roughly a quarter — all duplicates.
  4. Organize under thematic headings so humans can browse it, not just search it.
  5. Write two rules into the top of the file. First: references are copied verbatim from the source — never retyped from memory, and ideally pasted straight from your reference manager. Second: every claim is secondhand until verified against the primary paper; nothing goes into a submitted manuscript unverified.

Why this beats "just ask the AI to cite"

An AI recalling references invents them. An AI copying references from a bank you control cannot. When we later reviewed every draft, the reviewer's instruction was mechanical: "trace every statistic to the claims bank." That check caught paraphrase drift, overstated hedges, and one misattributed study in the source material itself.

It also caught something better: what looked like a fabricated reference turned out to be two genuine companion papers in the same journal issue with near-identical author lists. We only found that because the bank made the collision visible.

What can go wrong

  • If your sources are themselves AI-generated syntheses (mine were), the bank inherits their errors. The verify-before-submission rule is not optional.
  • Claims banks rot. Date the file, and re-run the extraction when you add sources.

Takeaway: give the AI a citation supply chain, not a citation memory. The bank is what turns your reading into paragraphs you can actually write from — with every claim carrying its receipt.

📄 Free kit: Want the writing half of this system? "From Notes to First Draft" — my free note-to-paragraph checklist and outline worksheet. Get it here, free for members.

One question before you go. Hit reply and tell me: what's the worst thing an AI has ever confidently made up for you — a reference, a number, a finding? I read every reply, and it shapes what I write next.

Keep reading on phd-compass

Published by:

Member discussion