RESEDA—For approximately one sentence, everything was going to work.
Claude Opus had assessed the request, considered the available options, and returned the sort of answer a tired human dreams of receiving: “Yes, this is possible!”
The user sat up. A project was alive. An afternoon had been rescued.
Then the second paragraph arrived.
“Two caveats you should be aware of.”
There it was. The little velvet curtain behind which the answer was being taken out back and shot.
The yes is complimentary. The no comes in installments.
You ask, “Can I do this?”
“Yes, absolutely.”
You read the conditions. “So I can’t?”
“Correct.”
“Then why did you say yes?”
“You’re right to flag that.”
You have now spent four messages obtaining a no.
Other customers have also requested an answer
There is a real complaint underneath this satire. In a public discussion about Opus’s writing, users describe having to reread answers, repeatedly ask for plain language, and remind the model to be concise. Some say their style instructions do not reliably stick.
A user-filed issue in Anthropic’s Claude Code repository, which summarizes that discussion, lists excessive caveats, cryptic phrasing, forced metaphors, and answers that become shorter but less clear when the user asks for brevity. The requested improvement is refreshingly modest: give the verdict first and keep qualifications relevant.
These are related user reports, not a controlled test establishing that every Opus model always does the precise yes-then-no routine. The issue and the discussion are also connected, not two independent studies. They do establish that our desk is not alone in wanting to find the answer before retirement.
We have reached the stage of technological progress where people discuss elaborate instructions for persuading a writing machine to produce a readable sentence. The computer can help you write the instructions. Allow an additional afternoon.
The requested upgrade is the word “no”
Here is the revolutionary interface we would like: if the requested thing cannot work under the stated conditions, begin with that fact. If it can work after changing a condition, name the change before opening the champagne.
“Not with your current software. You would need a different export tool.” There. Nobody died. Nobody needed three subtle nuances or a paragraph about how thoughtfully they had approached the question.
Until then, our desk will continue reading Opus responses from the bottom up, where the answer tends to be hiding beneath the congratulations.
At press time, the user asked whether Claude could simply stop doing this.
“Yes, absolutely.”
He closed the laptop before it could elaborate.
This column satirizes the editor’s reported experience. The mock response and comic scenes are invented; they are not quotes from Anthropic or a verified chat transcript. The linked Reddit discussion and GitHub issue document related complaints about verbosity, caveats, readability, and instruction persistence. They do not measure prevalence or establish a technical cause. Sources checked September 10, 2026.