THU, SEPTEMBER 24, 2026
Independent · In‑Depth · Practitioner‑Tested
Claude Productivity

Wide-Sweep Research Prompts: The 200,000 to 20 Funnel

Anthropic ran roughly 950 agents for 21 hours to narrow 200,000 reverse transcriptases to 3,500 candidates and then to 20, and human scientists took it from there. The transferable part is not the biology - it is the shape. Agents do the wide, consistent, high-volume filtering pass; people do the shortlist. These seven prompts run that funnel on any large search space, whether that is papers, patents, competitors, log lines or job listings.

⌨️ 7 prompts 🕐 Updated Sep 23, 2026
💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.
1
Write the Filter Before You Search
Criteria written after you start looking get bent by what you find. Write them first.
I am about to search a large space for [describe what you are looking for].<br/><br/>Before any searching happens, write the inclusion and exclusion criteria I will apply to every candidate.<br/><br/>Rules:<br/>1. Every criterion must be checkable from the item itself, not from judgement<br/>2. State what makes something a NO as explicitly as what makes it a YES<br/>3. Include an UNSURE category and define exactly what lands there<br/><br/>Output the criteria as a numbered list. Do not search yet.
2
The Wide Pass
The consistent, boring pass across everything. Run it in batches and concatenate.
Apply these criteria to every item in the batch below.<br/><br/>CRITERIA:<br/>[paste the criteria from the previous prompt]<br/><br/>For each item output one line only: item_id | YES / NO / UNSURE | the single criterion that decided it<br/><br/>No commentary. No summary. Do not skip items. Do not reorder them.<br/><br/>If an item is ambiguous, mark UNSURE rather than guessing - the UNSURE pile is the point.<br/><br/>BATCH:<br/>[paste]
3
Audit the Filter
Catches a bad filter before you have run it across 200,000 items and trusted the output.
Here are 30 items I marked NO and 30 I marked YES in the wide pass.<br/><br/>Read them and tell me where my criteria are doing the wrong thing. Specifically:<br/>1. Any NO that looks like it should be a YES<br/>2. Any criterion that is firing far more often than the others<br/>3. Any pattern in the UNSURE pile suggesting a missing criterion<br/><br/>Do not re-classify anything. Critique the criteria, not the items.<br/><br/>SAMPLES:<br/>[paste]
4
Rank the Survivors
3,500 to 20. Scoring on named dimensions beats asking for "the best ones".
These items all passed the filter. Rank them.<br/><br/>Score each one from 1 to 5 on: [dimension 1], [dimension 2], [dimension 3].<br/><br/>Output a table: item_id | score 1 | score 2 | score 3 | total | one-line reason<br/><br/>Sort by total, descending.<br/><br/>Do not editorialise beyond the one-line reason, and do not recommend anything - I am choosing, not you.<br/><br/>ITEMS:<br/>[paste]
5
Steelman the Top Five
The cheapest step in any research funnel is finding out early which finalist is wrong.
For each of the top five items, write the strongest case that it is NOT worth pursuing.<br/><br/>Be specific: what would have to be true for this to be a dead end, and how would I find that out cheaply?<br/><br/>One paragraph each. Do not balance it with the positive case - I have already got that.<br/><br/>ITEMS:<br/>[paste]
6
Design the Verification
Anthropic's result counted because humans put the protein in a cell. Design your wet lab.
For the item below, design the smallest experiment, query or check that would tell me whether the finding is real.<br/><br/>Constraints:<br/>1. It must be cheap and fast relative to fully pursuing it<br/>2. It must be able to return a clear negative, not just a positive<br/>3. State in advance what result would make me abandon this<br/><br/>If no cheap check exists, say so rather than inventing one.<br/><br/>ITEM:<br/>[paste]
7
Write Up What You Do Not Know
Anthropic led with "we don't yet know its function". That is why the result was believed.
Draft the findings summary for this search.<br/><br/>Structure: what was searched and how much of it, what criteria were applied, what survived, what was verified, and - as its own section with a heading - what remains unknown.<br/><br/>The unknown section must be specific. "Further work is needed" is not a finding. Name the exact questions that are still open and what would answer them.<br/><br/>Do not overstate. If the function or mechanism is unknown, that sentence goes in the summary, not the appendix.<br/><br/>MATERIAL:<br/>[paste]