AI's tidy answer to "who's waiting on whom?" can blame the wrong person. Make it quote its source.
What the Reel says
AI's tidy answer to "who's waiting on whom?" can blame the wrong person. Make it quote its source.
Long threads hide who's stuck. I tested 12 made-up emails about a café's order.
The first answer said three people waited on the designer to check a color. Only one had asked.
It dropped the deadline: approve the cup design by 10 November, or cups may miss the reopening.
So the fix: tell it to copy the exact line behind each row first.
This time it kept the deadline, copied word for word.
One wrong row gave itself away: its quote was a question the next email settled.
Next time, ask for its sources, then Ctrl+F each one before trusting the table.
Sources and assumptions
- Anthropic: Prompting best practices (Claude docs): Exact wording: 'Ground responses in quotes: For long document tasks, ask Claude to quote relevant parts of the documents first before carrying out its task. This helps Claude focus on the relevant content and ignore the rest of the document.' The page also gives an example prompt that puts quotes in <quotes> tags first and then answers from them. (checked 2026-11-08)
- Our test runs: GPT-5.5 via the OpenAI API (engine/ai-test.js), record.tests ids before-gpt and after-gpt: The real outputs quoted in this record. Both prompts were run on 2026-11-09 UTC (evening of 2026-11-08 Pacific) on the same fictional 12-message thread. (checked 2026-11-08)
Assumptions:
- The email thread is fictional: Bright Cup Cafe (Maya), Northline Packaging (Leo, sales; Dana, print studio) and a freelance designer, Sam. Prices and dates are made up and refer to no real company.
- The tests ran in GPT-5.5 through the OpenAI API (label 'GPT-5.5 (OpenAI API)', Nov 2026), not the ChatGPT app. The Claude runs ('before' and 'after' in tests.json) failed to run inside this research session and have no output. The post must not show or claim any Claude result.
- The technique source is Anthropic's documentation for Claude. The post may say it's Anthropic's documented tip and that it also worked in our GPT-5.5 test, but not that OpenAI recommends it.
- Who's correct after message 12, judged by hand: Maya is waiting on Dana for a revised proof with the logo 10% bigger (msg 10). Maya is waiting on Sam to check the green (msg 10). Northline/Leo is waiting on Maya's approval by 10 November (msgs 7, 9). Maya is waiting on Northline finance to confirm the deposit landed (msg 12). Already answered: the price (msg 2), the logo file (msgs 4, 8), the outlines (msg 8), the proof by Thursday (sent msg 9), 'Is that doable?' (answered msg 7), and the deposit request (paid msg 11).
- Every quote in the after-gpt <quotes> block and table was checked by hand and matches the thread word for word.
- The 'cups by 21 November' row in after-gpt counts as the bad row: its quote is Maya's msg 6 question, which Leo answered in msg 7 ('21 November works if the proof is approved by 10 November'), and the prompt said to leave out anything a later message answered. You could argue Maya is still waiting for the cups, so the post should call it 'a question already answered in message 7', not an invented row.
- Privacy line for the post: remove names, account numbers and other personal details before pasting a real thread, and check the tool's data policy or your employer's approved tools first.


