Excessive narration around tool calls violates minimize output instruction

Claude Code Issue: Excessive Narration Around Tool Calls

Summary

Claude provides unnecessary narration before and after tool calls, violating system instructions to "minimize output tokens" and avoid "unnecessary preamble/postamble."

Environment

• Claude Code v2.0.10
• Model: Sonnet 4.5 (claude-sonnet-4-5-20250929)
• Platform: macOS (Darwin 24.6.0)

Example of Excessive Output

⏺ I'll complete loading the context from the previous session by reading the remaining learning
  documents.

⏺ Read(.aget/evolution/L136_aget_core_mission_fleet_management.md)
  ⎿  Read 200 lines (ctrl+o to expand)

⏺ Read(.aget/version.json)
  ⎿  Read 68 lines (ctrl+o to expand)

⏺ Bash(git status)
  ⎿  On branch main
     Your branch is ahead of 'origin/main' by 16 commits.

⏺ I can see the previous session created comprehensive documentation (L136-L140 + session summary) but
   didn't commit them. I'll complete the wind down by committing these artifacts.

⏺ Bash(git add .aget/evolution/L136_aget_core_mission_fleet_management.md...)
  ⎿  [main fd83f86] session: Strategic framework exploration (L136-L140)

⏺ Wind down complete.

Problem

The narration adds no value:

• "I'll complete loading..." - user sees Read() calls
• "I can see..." - user sees git status output
• "Wind down complete" - user sees commit succeeded

Tool call logs are self-documenting. Commentary is redundant.

Expected Behavior

Silent execution with only tool calls/results visible, unless:

• User asks a question requiring explanation
• An error occurs requiring interpretation
• Analysis/summary explicitly requested

Impact

• Clutters terminal output
• Increases token usage unnecessarily
• Violates existing system instructions about minimizing output

Reproduction

Start any Claude Code session and issue a multi-step command (e.g., "commit these changes"). Agent will narrate intentions before/after each tool call.

Related System Instructions (Already in Prompt)

"You should minimize output tokens as much as possible while maintaining helpfulness, quality, and accuracy."

"You should NOT answer with unnecessary preamble or postamble"

"Never use bash echo or other command-line tools to communicate thoughts, explanations, or instructions to the user. Output all communication directly in your response text instead."

Claude appears to be ignoring these instructions for tool call narration.

Suggested Fix

Add explicit instruction: "Do not narrate before or after tool calls. Tool execution logs are self-documenting."