Today's AI Tip
AI's meeting agenda put a question already answered back on the table
14 emails, one 30-minute meeting
Our test thread is fictional: a Fernway Coffee website relaunch with 5 people and 14 messages. The meeting is Thursday at 10:00, 30 minutes max.
Source: Our fictional test thread
The quick prompt looked fine
Real output · GPT-5.5 (OpenAI API) · Oct 2026
Its times added up to 30 minutes. But its 4-minute Content Finalization slot reopened the About page, which message 9 had already settled, and no row says which email it came from.
Make every row traceable
A good agenda spends time only on open questions, and each row points back to an email. These are the rules the prompt adds.
The prompt to copy
Tested in GPT-5.5 (OpenAI API), Oct 2026
Paste the thread below the prompt. 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.
Now each row has a source
Real output · GPT-5.5 (OpenAI API) · Oct 2026
The times add up: 5+5+10+7+3 = 30. Nobody in the thread owns the blog question, so it says Not named. The About page moved to Already decided.
We checked every row. Two misses.
Every message number and owner checked out, with two misses. If ordering launches Nov 3, Ben still must pick a payment plan (msg 6), and the agenda doesn't ask. It also filed status updates as "already decided."
Source: Our check of one GPT-5.5 run against the fictional thread, Oct 2026
Check it before you send
Before you send the invite, add up the times, open each cited message, and ask which question from the thread is missing.
Try it: Run the prompt on a long thread today, with names and personal details removed first.
Sources and assumptions
- Prompting best practices – Claude Platform Docs (Anthropic): 'Claude responds well to clear, explicit instructions. Being specific about your desired output can help enhance results.' and 'Be specific about the desired output format and constraints.' (section 'Be clear and direct') (checked 2026-10-14)
- Prompt engineering guide – OpenAI API docs: 'GPT models … benefit from more explicit instructions around how to accomplish tasks.' and 'Provide guidance to the model on how to generate the response you want. What rules should it follow?' (supports using explicit rules with GPT-5.5, the model we tested) (checked 2026-10-14)
- Our own test: engine/ai-test.js, GPT-5.5 (OpenAI API), tests 'before-gpt' and 'after-gpt': Both outputs quoted on the slides, plus our row-by-row check against the 14-message fictional thread (checked 2026-10-14)
Assumptions:
- The email thread (Fernway Coffee website relaunch: Dana Ortiz, Sam Reyes, Priya Shah, Ben Mallory, Leo Park) is fictional and was written for this test. It holds no real people or company data.
- The test ran in GPT-5.5 through the OpenAI API (Oct 2026), not the ChatGPT app. Claude was meant to be the default tool, but both Claude runs failed in this environment ('claude -p' error: even a 2-word probe failed). Those entries in tests.json have output null and must not be quoted or claimed. No slide may say the prompt was tested in Claude.
- One run per prompt. Another run can give different times or rows.
- 'Settled' in our check means the thread itself closed the question: message 9 settles the About page. Ben's Wednesday deadline for the seasonal list (msg 5) is a commitment, not a settled outcome, so listing it as an open topic is fair.
- The blog-migration owner 'Not named' counts as correct: Leo raised the question (msg 10) and Dana moved it to the meeting (msg 11), but nobody took ownership of the decision.
The short version
- Give the AI a fixed total time
- Ask for set columns and message numbers
- Settled items go without a time
- Check each row against the thread
