When AI summarizes a meeting, have it quote every open concern exactly.
What the Reel says
When AI summarizes a meeting, have it quote every open concern exactly.
I asked GPT-5.5 to sum up a made-up team call in three bullet points.
The plan was to switch to a new courier right before the busiest two weeks.
Luis, the warehouse lead, pointed out they'd never used its pickup system.
The AI's version just said “despite Luis noting some risk.” What the risk was got cut.
So I added one line: list every unresolved concern word for word, with the speaker's name.
This time, everything he said came back in a separate list. The bullets still claimed the team agreed.
In my test, the doubt only survived in the quotes, so read those first.
Sources and assumptions
- Reduce hallucinations – Claude Platform Docs (Anthropic): Anthropic recommends asking the model to extract word-for-word quotes to ground its answer in the actual text, and to cite quotes for its claims so the response can be checked. Note: the page aims the quote-first advice at long documents (over 20k tokens); a short meeting transcript is below that, so the post must not say the docs target this case. It also says these techniques reduce mistakes but don't eliminate them. (checked 2026-11-04)
- Real test runs, GPT-5.5 (OpenAI API), Nov 2026 – record.tests: Test outputs ops-plain, ops-plain-2 and ops-quotes were run with engine/ai-test.js on 2026-11-05 UTC (2026-11-04 Pacific). The URL is only the model's documentation; the outputs live in record.tests. (checked 2026-11-04)
Assumptions:
- The transcripts are fictional (Jen, Luis, Amara, Tom on an ops call; Dana, Marco, Priya, Sam on a menu launch), written for the test. No real meeting data.
- The tests ran on GPT-5.5 through the OpenAI API, not the ChatGPT app and not Claude. The Claude runner (claude -p) failed in this environment, and the entries 'before', 'after' and 'probe' in tests.json hold null outputs from those failed attempts. Don't use them, and don't claim a Claude result.
- Results vary run to run. 'despite Luis noting some risk' / 'despite Luis flagging some risk' appeared in both plain runs (ops-plain, ops-plain-2).
- Limitation 1: in a shorter meeting where the concern was stated plainly (spring menu tests 'plain', 'short-plain', 'short-plain-2'), GPT-5.5's plain summary did keep Priya's concern and its reason. The loss showed on the softer, politely dropped objection. The post must say 'in my test' and not claim AI always drops objections.
- Limitation 2: even with the fix (ops-quotes), the summary bullets still said 'The team agreed to switch to the new courier', although Luis had said 'But it is your call.' The quoted list is what keeps the objection, so read it, not just the bullets.
- Privacy: speaker quotes need names. Only paste transcripts your workplace lets you share with that tool, or swap names for roles (Warehouse lead, Finance) first, and check the tool's data policy or your employer's approved tools before uploading a real transcript.
- The quote-first docs cited are Anthropic's, written for Claude and long documents. They're the background for the technique, not proof that it works on GPT-5.5; the GPT-5.5 proof is the test.


