The internet has one message for Opus 5. SHUT UP!

The smartest model Anthropic has shipped, and thousands of people want it to stop talking.

TL;DR:
Use custom output styles to wrangle opus 5 verbosity to a level you like.
👉 Here is the full prompt set.

Hey Adopter,

Opus 5 landed a few weeks ago. By every benchmark that matters, it is the strongest model Anthropic has shipped. Talk to it in plain conversation, though, and something else happens.

Ask it a plain question and the reply arrives wrapped in paragraphs of throat clearing, packed with terms it half invents. You finish reading more tired than when you started. Reddit split into camps within the first couple of weeks and coined its own words for it. “Claudeslop.” “Benchslop,” for the gap between the benchmark score and how the model actually reads.

“It’s too talkative.”

That was the top-voted answer to whether the launch was hype or the real thing. One reader, quoted in a widely read review of the launch, said his eyes now glaze over reading Claude’s replies, and that human writing in the same register has started doing the same thing to him. The model is not wrong more often. It is just exhausting to decode.

So what’s the issue?

Anthropic told us why themselves. Thariq, who works on Claude Code, posted that the team had removed more than 80 percent of Claude Code’s system prompt for models like Opus 5, with no measurable drop on their coding evaluations.

“We were over-constraining Claude Code.”

Their own transcripts showed the old system prompt fighting itself. One rule said add documentation. Another said do not comment. Multiply that across a year of accumulated skill files and CLAUDE.md rules, and the model was spending part of every reply untangling contradictions instead of doing the work. So they cut the rules and let the model use its own judgement.

Where it helps and where it doesn’t

Good trade for a coding benchmark. Anthropic never said this is also why the conversational voice got denser, but fewer guardrails almost certainly means more of the model’s own instincts show through unfiltered, including its instinct to write for another version of itself, not for you.

---

Here’s what you can do about it.

Do not fight that instinct every message. Decide the rules once, the way Anthropic did. Claude Code has a feature for exactly this, called output styles. One markdown file, added to the system prompt for the whole session, and every reply follows it from then on.

You can just copy-paste the following or use any of the prompts in the prompt pack inside Claude Code and have a conversation about it, and it will create those styles for you

---
name: no-slop
description: Direct human voice. Zero AI filler patterns.
keep-coding-instructions: true
---
Speak like a precise senior engineer who is tired of corporate writing.

Lead with the point. Specific nouns and verbs. Cut every sentence that does not change what the reader will do or understand. No "it's worth noting," rule-of-three padding, or vague intensifiers. Opinion is preferred when a recommendation is required. Code stays exact.

Save that as no-slop.md in .claude/output-styles, then run /output-style and pick it. I built six of these for different situations, one for when you are running on no sleep, one for executive updates that answer first and explain after, plus two prompts that turn a bad reply you paste in into a brand new style file. The full set is in the Shut up Opus prompt pack on Right Click Prompt.

Opus 5 is not getting simpler. The layer between you and its default voice is still yours to set. Spend the two minutes today, not the next time a reply wears you out.

Adapt & Create,
Kamil