Don't ask AI to summarize the memo. Ask for these four lists.
The default: hit summarize
Real output · Claude Sonnet 5.5 · Sept 2026
What you get if you just say 'Summarize this': accurate prose to reread, with no page numbers, and gaps mentioned only in passing.
Ask a different question
A summary answers 'what does it say?' You need 'what do I have to act on, by when, and who owns it?'
The four-list prompt
Tested in Claude Sonnet 5.5 and GPT-5.5 (OpenAI API), Sept 2026
Paste your document, then this prompt.
Four tasks, no owner
Real output · Claude Sonnet 5.5 · Sept 2026
The same prompt, run on the same memo, flagged four ownerless tasks instead of one, each with a page number. One of them didn't hold up (next slide).
Source: Our own AI test, Claude Sonnet 5.5, Sept 2026
Check every 'not found'
Claude marked the deposit decision not found, but page 3 says the partners decide it Oct 6. GPT-5.5 caught only one ownerless task total.
Try it: Open the page next to any 'not found' before you trust it.
Source: Our own AI test vs. source memo, Sept 2026
Use it, then verify
Check each flagged page. Page numbers need visible page breaks — ours were typed in. Remove names, account numbers and other personal details first, and check your employer's approved tools before uploading real documents.
Try it: Paste your next long memo or meeting pack with this prompt; type [Page 1], [Page 2] markers in if the paste lost them.
Sources and assumptions
- Our own test run, engine/ai-test.js, Claude Sonnet 5.5 via claude -p: The before ('Summarize this document.') and after (four-list prompt) outputs on the fictional Harbor Lane Bakery memo, recorded in record.tests with ids 'before' and 'after' (checked 2026-09-29)
- Our own test run, engine/ai-test.js, GPT-5.5 via the OpenAI API: The same four-list prompt run in GPT-5.5 (record.tests id 'after-gpt'): one 'not found' owner, and Marco's before-Oct-6 quote missing from its Deadlines list (checked 2026-09-29)
Assumptions:
- The sample document is a fictional 3-page planning memo written for this test, with [Page 1]/[Page 2]/[Page 3] markers typed into the text. Real PDFs pasted as text may not carry page markers, so the page references depend on the tool seeing page breaks (uploaded PDFs, or markers added by hand).
- Each prompt was run once. Outputs vary between runs, so a reader's result will differ in wording and may differ in which gaps it flags.
- Claude was run through claude -p with a plain assistant system prompt (Claude Sonnet 5.5), and GPT-5.5 through the OpenAI API, not the ChatGPT app.
- The web sources for general prompting guidance could not be fetched in this session, so the post claims nothing beyond what our own tests showed.
Save this template
- Ask for four lists, not a summary.
- Every item gets a page number.
- Not found still needs a human check.
- Different models catch different gaps.


