“Not Reported” Is Not “Not There” — And All Ten Prompts Say So

The card itself is almost bare. “NEW: Use Claude as your research assistant.” Three pills underneath — literature review, knowledge map, research gaps — and a promise to turn forty-plus papers into all three. The ten prompts are in the comments, written out in full, and they are the most carefully built set this page has published.

You can tell within one sentence of each. These are not “write me a literature review”. They specify paper IDs, source locations, effect sizes, verdict columns, comparison groups. Somebody who has actually done a systematic review wrote these.

But the thing worth noticing only shows up when you read all ten together and look at what each one forbids.

Ten prompts, ten guards, one mistake

Every single one of the ten carries a clause whose only job is to stop the model turning “I do not have this” into “this does not exist.” Read them in a row:

• 1. Knowledge map — “Label every connection as an observed association, tested causal effect, theoretical proposal, or inference… Exclude unsupported connections.”
• 2. Evidence database — “Mark missing information as ‘not reported.’”
• 3. Field timeline — “Never label a study ‘the first’ without verification.”
• 4. Review audit — a verdict column with four values, one of which is “unverifiable”, plus “flag anything that cannot be confirmed.”
• 5. Research gaps — “Label findings as ‘gaps within this collection’ unless a broader search verifies them. Never assume that an unaddressed topic has never been researched.”
• 6. Contradictions — “Separate evidence-backed explanations from plausible hypotheses.”
• 7. Paper search — “Report the actual number of relevant papers found. Never fabricate sources or pad the results.”
• 8. Study design — “Do not invent sample-size calculations; identify the information required to calculate one.”
• 9. Evidence strength — “Distinguish missing reporting from confirmed methodological flaws.”
• 10. Literature review — “a reference list containing only sources actually used.”

Ten for ten. Nobody writes ten of those by accident, and they are not generic “don’t hallucinate” boilerplate. Each one names a specific gap and a specific wrong way to fill it.

Diagram: eight pairs a model cannot tell apart

Line the clauses up and they resolve into eight distinctions. On the left is the thing a model will produce if you say nothing; on the right is what the clause forces. In every case the two outputs look identical on the page — that is the whole problem.

There is a reason this has to be said out loud, and it is not that the model is careless. A language model completes text. A gap in a table is, structurally, a place where text should go. Nothing in that process distinguishes “the paper does not report this” from “I did not find where the paper reports this” from “no paper reports this” — those are three completely different facts about the world that arrive as the same blank cell.

So the author’s clauses are doing something more precise than discouraging invention. They are giving the model somewhere to put ignorance. “Not reported” is a value. “Unverifiable” is a verdict. “Gaps within this collection” is a scope label. Each one creates a legal output for I don’t know, and a model that has a legal way to say it will use it. A model with no such option fills the gap with the most plausible-looking thing, because that is the only move available.

That also explains why these prompts avoid the usual problem with negative instructions. Anthropic’s own prompting guidance says to tell Claude what to do rather than what not to do — and on this card, nine of the ten prohibitions arrive attached to a positive category to use instead. “Never fabricate sources” would be weak on its own; “report the actual number of relevant papers found” is the instruction that makes it work.

All ten, and the order to run them in

The card presents them as a list. They are not a list — they are a pipeline, and two of them only work if another has already run.

Two things stand out once they are in this order.

Prompt 2 is the one that matters. Five of the other nine read from the evidence table it builds. If that table is wrong — a mis-transcribed sample size, a fabricated effect size — the error propagates into the map, the timeline, the contradiction analysis, the appraisal and the review, and none of those five will catch it, because they are reading the table rather than the papers. Build that one slowly.

Prompt 4 is the only one that judges. Nine prompts produce something; one checks it. And it is the only one in the set that compares the output back against the sources rather than against the previous step — which is exactly the property that makes it worth running. We have written before about why the thing that makes the output cannot be the thing that grades it; here, at minimum, run prompt 4 in a fresh conversation that has not seen your draft being written.

One more detail worth stealing: nearly every prompt carries a size. Fifteen nodes. Five gaps. Three designs. A 200-word synthesis. A 1,200-word review. Up to forty papers. That is the same discipline we counted on a very different deck last week, and it does the same job here: it stops the model padding.

The arithmetic of forty papers

“Turn 40+ research papers into…” is the headline promise, and it is the one claim on the card with a number in it. So here is the number behind the number.

None of that is an argument against the deck. It is the practical shape of the job: build the evidence table in batches, then run the other nine prompts against the table rather than against the papers. That is cheaper, it is faster, and it is what prompt 2 is quietly telling you to do. The cost of ignoring it is paying half a million tokens nine times over.

The one check that is not on the card

Given how carefully these ten were written, the gap is conspicuous. Not one of the ten prompts asks whether a paper has been retracted.

Retraction is the single most consequential fact about a paper, it is binary, and it is checkable. A retracted study still sits in PDF form on your hard drive looking exactly like a good one — the file does not change. And a model reading that PDF has no way to know, because the retraction notice lives somewhere else entirely.

Add a column. In the evidence table from prompt 2, next to author and year, put retraction status, and set its default to “not checked” rather than blank — which is the same trick the whole deck already uses, applied to one more thing. Then check it yourself against the publisher’s page or a retraction database, because that is an outside-the-corpus fact and no prompt can reach it.

The same applies, slightly less sharply, to funding source and declared conflicts of interest. Prompt 9 appraises sampling, confounding, measurement and follow-up — a proper risk-of-bias list — but who paid for the study is not on it, and in several fields that is the first thing a reviewer looks at.

Both of these are the same category of problem the deck is already built around: a fact the papers in front of you cannot supply. The author clearly understands that category. These two just did not make the ten.

Four honest caveats

1. We did not run these on a real corpus. This is a reading of ten prompts and some arithmetic, not a trial. We cannot tell you how well prompt 2 actually extracts effect sizes from forty PDFs — only that its guard clauses are the right ones and that it is the step everything else depends on.
2. The token figures are an estimate with a range. Article length varies by field by a factor of three, and a PDF full of tables and figures tokenises differently from clean text. Treat 320,000–530,000 as an order of magnitude, not a quote.
3. Prices and windows move. The context-window and pricing lines were read from Anthropic’s own pricing page in October 2026, and quoted for the cheapest current model so the comparison is as favourable to the card as possible. Check before you budget.
4. We are agreeing with the card, not correcting it. Its footer already says Claude can invent sources and that every citation needs checking against the original. Our addition is one column — retraction status — and one observation: that the ten prompts are a pipeline, not a menu.

The one-line version

These ten prompts are built around a single idea, and it is the right one: a model cannot tell the difference between “the paper does not say” and “I did not find where it says it” and “nobody has ever said it” — so you have to give it somewhere to put each one. “Not reported.” “Unverifiable.” “Gaps within this collection.” Ten prompts, ten legal ways to say I don’t know, and that is what makes them usable for research rather than for a blog post.

Steal the technique even if you never review a paper. Whenever you ask for a table, name the string that goes in an empty cell. We have made the related point before about prompts that quietly ask for facts the model does not have — this deck is what the other half of that lesson looks like when someone gets it right.