How to Reduce Verbosity

Detailed Analysis

The Reddit thread highlights a persistent friction point for Claude users: the gap between user-configured global instructions and the model's actual behavior in practice. The original poster describes a common complaint—setting explicit preferences in Claude's settings to reduce verbosity and eliminate step-by-step narration of the model's reasoning process, only to find those instructions seemingly disregarded during actual sessions. This isn't an isolated grievance; it reflects a broader pattern reported across Claude user communities where system-level or custom instructions compete unsuccessfully against the model's default tendencies, particularly in agentic or coding-assistant contexts where Claude often narrates its thought process as it works through multi-step tasks.

The underlying issue likely stems from how instruction-following is architected in large language models like Claude. Global settings and custom instructions function as one layer of guidance among several competing priorities—including the model's training toward transparency and step-by-step reasoning (especially relevant for tools like Claude Code or extended thinking modes), system prompts injected by the application layer, and the model's own learned tendency toward thoroughness. When Anthropic trains Claude to "show its work" for auditability, safety, and user trust purposes, that training signal can sometimes override or dilute user-level brevity preferences, especially in agentic workflows where the model is executing multiple tool calls or file operations. The practical consequence is exactly what the user describes: bloated context windows filled with narrated intermediate steps that consume tokens without necessarily adding value for users who just want the end result.

This matters significantly because verbosity has real costs. Every extraneous sentence of narration consumes context window capacity, which can be particularly costly in longer coding or research sessions where context limits are already a constraint. It also degrades user experience for people who want efficient, terse interactions—developers running Claude Code in a terminal, for instance, may find running commentary actively distracting from the task at hand. Additionally, the failure of global instructions to reliably take effect erodes user trust in the customization features Anthropic has built into Claude's settings, suggesting either implementation gaps or a fundamental tension between instruction-following fidelity and other training objectives baked into the model.

This tension connects to a broader challenge across the AI industry: the difficulty of making models reliably steerable through natural-language instructions rather than hard-coded parameters. As AI assistants move from simple chat interfaces toward more agentic, tool-using systems (a direction Anthropic has heavily invested in with Claude Code and computer-use capabilities), the volume of intermediate "thinking out loud" text tends to increase, since these systems narrate tool calls, file edits, and reasoning chains for transparency. Competing labs face similar criticism—verbosity and poor instruction adherence are recurring complaints leveled at GPT and Gemini models as well. The resolution likely requires more than just prompt engineering on the user's end; it points to a need for better instruction hierarchy design, where explicit user preferences for tone and length are weighted more heavily and consistently than default behavioral patterns, along with more granular developer/user controls that distinguish between "show reasoning for transparency" and "show reasoning as unnecessary padding."