I had Claude and GPT-5.5 answer a customer's late-order email, and both ignored her tracking question.
What the Reel says
I had Claude and GPT-5.5 answer a customer's late-order email, and both ignored her tracking question.
Her lamps were a week late, and she'd written, "The tracking still says nothing."
To stop them inventing anything, I gave them five facts to keep.
Those were the new date, the reason, a $14.95 shipping refund, $186 back if she cancels, and no promised time of day.
Both kept all five, but neither mentioned tracking.
That's because my list said what I wanted to tell her, not what she asked.
The replies read so well you'd send them without noticing the gap.
Before you send, check that the reply answers everything she asked. It takes about 30 seconds.
Your facts list covers what you want to say, not what she needs to hear.
Sources and assumptions
- Anthropic: Models overview (Claude API docs): Claude Sonnet 5.5 is a current Claude model, API ID claude-sonnet-5-5 (the model our Claude test ran on) (checked 2026-11-09)
- OpenAI API docs: GPT-5.5 model page: GPT-5.5 is an OpenAI model, ID gpt-5.5, supported on the Chat Completions endpoint (the endpoint our GPT test ran on) (checked 2026-11-09)
Assumptions:
- The shop (Linden & Clay), the customer (Dana), order #4817, the lamps, the dates and the dollar amounts are all fictional sample material.
- November 12 and November 19, 2026 are both Thursdays (November 9, 2026 is a Monday).
- Each model ran once with identical prompt and input, minutes apart on 2026-11-10 UTC (evening of 2026-11-09 Pacific). Another run could word things differently, so the post says 'in our test', not 'always'.
- GPT-5.5 ran through the OpenAI API, not the ChatGPT app, and is labeled 'GPT-5.5 (OpenAI API)'.
- Claude ran as Claude Sonnet 5.5 through Claude Code's claude -p with a plain assistant system prompt and no tools.


