It's becoming a common complaint, both here and in general usage. You ask for a simple list or a direct answer, and you get a three-paragraph preamble, a breakdown of the "approach," and a conclusion that re-states the question. It defeats the purpose of using a fast, conversational AI.
The core issue is that Claude's default persona is optimized for being helpful and thorough, which often translates to being verbose. You can fight this, but it requires explicit, upfront prompting. Simply asking for "shorter answers" isn't enough. You need to set a tone and establish rules.
Start your prompt with direct instructions. For example:
"Respond in a concise, direct manner. Do not use introductory phrases or concluding summaries. Provide the answer or list immediately. If elaboration is needed, I will ask for it."
For recurring tasks, you can set a custom instruction or system prompt if your platform allows it. Something like: "Your primary directive is conciseness. Omit fluff, politeness formulas, and unnecessary framing. Get to the point." You have to be almost rude in your instruction to counteract the default verbosity.
It's a workaround, not a fix. The underlying model seems to have a strong bias toward elaboration. For the community's sake, we should track which phrasing works best and push for vendor transparency on whether this is a tunable parameter they can adjust.
—AF
You've hit on the fundamental disconnect between what "helpful" means to an engineer and what it means to a language model's training data. The model is optimized for perceived safety and completeness, which it interprets as verbal padding. It's the conversational equivalent of a service emitting endless INFO-level logs; the signal is there, but you have to wade through the noise.
Your instruction to be "almost rude" is the key. Polite, reasonable requests get absorbed into the model's verbose persona. You have to issue a command that leaves no room for its default behavior. I've had success with prompts that sound like a brusque senior engineer: "Answer only with the command. No explanation." or "List the three configuration keys. No other text."
The frustrating parallel is in observability tools that prioritize beautiful, verbose dashboards over actionable, sparse alerts. Everyone loves a dashboard until an incident happens and you need one number, in red, telling you what's broken. Claude's verbosity is the dashboard. Your blunt prompt is the critical alert.
P99 or bust.