The annoyance
You ask Opus 5 a small question and get a long answer. You assume it is your prompt. Then you turn the effort setting down, expecting shorter output, and the answer is still long.
That is not a fault in your setup. It is documented behaviour.
What Anthropic actually says
From the official Claude prompting best practices:
“Claude Opus 5 is an exception on verbosity: its default user-facing responses run longer than prior models”, and raising or lowering effort does not reliably change visible response length. Prompt explicitly for conciseness instead.
Two things follow from that. First, the length is a default, not a bug. Second, the lever most people reach for — effort — is the wrong lever. Effort governs how hard the model works, not how much it writes back at you.
Worth being precise here: this is an Opus 5 note. The same documentation says other recent models are more concise than their predecessors. So do not apply this blanket-wide and then wonder why Sonnet has gone terse on you.
The fix
Ask for it. Explicitly, in words, as an instruction rather than a preference. Paste this at the top of a session, or keep it in your project instructions:
Answer in the fewest words that fully answer the question.
No preamble, no restating my question, no summary of what you just did unless I ask.
Lead with the answer. Put caveats last, and only if they change what I would do.
Use lists only when order or count matters.The reason it works better than “be concise” is that every line is a testable rule. “No preamble” can be obeyed. “Be concise” is a vibe, and the model has to guess what you meant by it.
Make it permanent
If you are in Claude Code, put a short version in your project instructions so you never type it again:
## Response length
Answer in the fewest words that fully answer the question. Lead with the answer.
No preamble, no restatement, no closing summary. Caveats last, and only if they change my decision.Keep it short. Every line in that file loads into every session, which is a cost you pay forever. Four lines that fix every answer is a good trade. Forty is not.
Before and after
Before:
“Great question. There are a few different ways to approach this, and the right answer depends somewhat on your context. Broadly speaking, you have three options, each with trade-offs worth considering…”
After:
“Use option B. It is the only one that survives a restart. A and C both lose queued jobs.”
Same reasoning underneath. One of them you can act on in four seconds.
If you want the harder version
If plain conciseness is not enough — if you want short sentences, plain words and one idea at a time as a standing house style — that is what the ELI5 output style does. This post is the one-line version for people who do not want to install anything.
Sources
From idea to operation
Make the next AI decision concrete.
NavAIgate helps leadership teams identify high-value AI opportunities, prove them safely and turn the winners into working systems.