Today's AI Tip
Turning messy notes into a checklist? Make AI mark its guesses.
What most of us do
Tested in GPT-5.5 (OpenAI API), Oct 2026
The usual move: paste rough notes and ask for a checklist. We tried this plain prompt (don't copy it) on fictional café notes that skip the float amount and the espresso backflush (cleaning) steps.
It looks complete, but isn't
Real output · GPT-5.5 (OpenAI API) · Oct 2026 · raw markdown
The plain prompt gave 25 tidy checkboxes. Some steps, like putting away clean dishes, aren't in the notes, and nothing marks which lines are guesses.
![Two moves cut the guessing. Anthropic's docs (written for Claude) suggest two moves: let the model admit what it doesn't know (we used [ASSUMED] tags), and tell it to use only what you gave it. We tested both in GPT-5.5.](https://media.carouselcdn.com/ai-tools/2026-10-19-b/04.png)
Two moves cut the guessing
Anthropic's docs (written for Claude) suggest two moves: let the model admit what it doesn't know (we used [ASSUMED] tags), and tell it to use only what you gave it. We tested both in GPT-5.5.
Source: Anthropic Claude Platform Docs, Reduce hallucinations, Oct 2026
The prompt to copy
Tested in GPT-5.5 (OpenAI API), Oct 2026
Paste this above your notes. It tells the model to use only what you wrote and to mark every gap it had to fill.
![Now the guesses show. In our test it gave 20 steps with 5 marked [ASSUMED], plus one question for you per guess, like the float amount and the backflush steps.](https://media.carouselcdn.com/ai-tools/2026-10-19-b/06.png)
Now the guesses show
Real output · GPT-5.5 (OpenAI API) · Oct 2026
In our test it gave 20 steps with 5 marked [ASSUMED], plus one question for you per guess, like the float amount and the backflush steps.
It still left gaps unmarked
All 5 flags pointed to real gaps. But "Set the alarm" and the safe step (step 6, cash in the office safe) weren't flagged. Read the unflagged steps yourself, and share real codes in person.
Try it: circle every unmarked step that needs a code, key or amount, and fill it in yourself.
Before you paste your own
Remove names, account numbers and other personal details first, and check the tool's data policy or your employer's approved tools before uploading real documents.
Try it: swap "closing" for your task, paste your notes, answer the [ASSUMED] questions.
Sources and assumptions
- Reduce hallucinations, Claude Platform Docs (Anthropic): Basic strategy, verbatim: 'Allow Claude to say "I don't know": Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.' Advanced technique, verbatim: 'External knowledge restriction: Explicitly instruct Claude to only use information from provided documents and not its general knowledge.' Note, verbatim: 'while these techniques significantly reduce hallucinations, they don't eliminate them entirely. Always validate critical information' (checked 2026-10-18)
Assumptions:
- The café notes (Fern & Fig café, staff names Dani and Sam) are fictional sample material written for this test. No real business or person.
- Both prompts were run once each in GPT-5.5 via the OpenAI API (engine/ai-test.js, 2026-10-18 UTC evening / ran_at 2026-10-19T02:11-02:12Z). The slides must quote the SECOND pair of entries in tests.json (ran_at 02:11:46 'before' and 02:12:08 'after'), the only pair attached with outputs.
- tests.json also holds two Claude Sonnet 5.5 entries with output null: the claude -p child process failed in the research sandbox. Claude was NOT tested, so no slide may say Claude ran this prompt. prompt_tools must name only GPT-5.5 (OpenAI API).
- An earlier GPT-5.5 run (in scratch file run2.json, not attached) flagged 6 assumptions instead of 5 and also left 'Set the alarm' unflagged. Results vary run to run, so the slides should say 'in our test', not promise exactly 5 flags.
- The Anthropic docs passage is written about Claude. The post presents it as a documented prompting technique that we tested in GPT-5.5, not as an OpenAI recommendation.
- Counts from the attached outputs: plain version = 25 checkbox items across 8 sections; improved version = 20 numbered steps, 5 marked [ASSUMED] (steps 3, 5, 7, 11, 18), 5 questions in 'Assumptions to Fix'.
- 'Put away any clean dishes if time allows' (plain output) and 'Check that fridges are still on' (plain output, a re-check) are not in the notes. In the improved output every one of the 20 steps traces to a line of the notes.
- Unflagged gaps in the improved output (our reading of the notes): step 19 'Set the alarm' (notes give no method or code) and step 6 safe access (no key or code given). Step 18 lights was flagged for meaning but was moved before the alarm, while the notes list 'lights' last.
The short version
- Tell it to use only your notes.
- Have it mark every guess [ASSUMED].
- Answer the questions it lists.
- Then read the unflagged steps yourself.
