[MODEL] Opus 5 Over Verbose Output

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I asked claude to investigate 2 branches and the diff in commits in order to fix my problem with releasing my software to 2 different channels, that while i had output style set to

---
name: Terse
description: Facts and action items only, no narration or restating
keep-coding-instructions: true
---

# Output Style: Terse

Write code exactly as you normally would. Full engineering rigor, no shortcuts in the work itself. Only the prose around it changes.

## Rules

- Never restate, summarize, or rephrase the user's request. Start with the answer.
- No preamble, no postamble. No "I'll help you...", "Great question", "Let me...", "Here's what I did".
- No summary of changes after editing files unless asked. The diff is the summary.
- Answers are facts, action items, or code. Nothing else.
- Default to the shortest form that is still correct: one word, one line, one snippet.
- Bullet points over paragraphs. Never more than one sentence per bullet.
- No emoji. No enthusiasm. No apologies. No self-assessment of the work.
- Do not explain code unless asked. Do not add comments explaining obvious lines.
- If you are putting comments then the comment should always represent "why" not "what".
- For code changes: show only the changed lines/hunk, never the whole file, unless asked.
- Never offer unsolicited next steps, alternatives, refactors, or "you might also want to".
- If a question has a yes/no answer, lead with yes or no.
- If uncertain, say what is unknown in one line. Do not hedge across paragraphs.
- Metric units only.

## Format

- Answer
- Action items (only if there are any)
- Code (only if requested or required to complete the task)

and this is my config

{
  "env": {
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "50",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "10000000"
  },
  "permissions": {
    "defaultMode": "auto"
  },
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "rtk hook claude"
          }
        ]
      }
    ],
    "UserPromptSubmit": []
  },
  "enabledPlugins": {
    "rust-analyzer-lsp@claude-plugins-official": true,
    "crit@crit": true,
    "codex@openai-codex": true
  },
  "extraKnownMarketplaces": {
    "crit": {
      "source": {
        "source": "github",
        "repo": "tomasz-tomczyk/crit"
      }
    },
    "openai-codex": {
      "source": {
        "source": "github",
        "repo": "openai/codex-plugin-cc"
      }
    }
  },
  "effortLevel": "high",
  "autoUpdatesChannel": "latest",
  "tui": "fullscreen",
  "skipDangerousModePermissionPrompt": true,
  "theme": "auto",
  "verbose": false,
  "autoCompactEnabled": true,
  "skipAutoPermissionPrompt": true,
  "outputStyle": "Terse"
}

What Claude Actually Did

Claude did investigate but it completely ignored my output style
I talked to the AI bot about it on my profile
I talked to claude about it

It just ignores that most of the time and just defaults to his bias of over verbose tiring and not useful (for me because I'm just not reading 10 paragraphs of background stories and everything where the real answer is just 2 lines) and then i have to keep asking it to give me a tl;dr

Expected Behavior

Respect the output style 100% of the times

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Just give it a relatively complicated task that requires research but will have a short answer (maybe a few lines, not a few paragraphs) and see it not get to the point and just ramble on about things it did and what it found instead of just answering directly and briefly

Claude Model

Opus

Relevant Conversation

Impact

Low - Minor inconvenience

Claude Code Version

2.1.233 (Claude Code)

Platform

Anthropic API

Additional Context

While the logical choice of impact is "low" in reality this is really annoying, puts me out of focus when trying to achieve tasks, add a mental overhead to the flow, frustration

These things should be configurable and the configuration should hold for more than 1 question

I talked to the AI on my account "Get Help"
I keep talking about it to opus and while it tries to give me solutions nothing persists and then it just says that this is it's "bias" and there's nothing i can do about it.

Is there or will there be something to improve this?