Boost Claude Code Output Quality and Speed by Tweaking 10 Hidden Settings

Effort Parameter

The default /effort was reduced from high to medium in March, causing the Agent to avoid deep reasoning. Temporarily you can run /effort high in a session, but the permanent fix is to set the environment variable:

export CLAUDE_CODE_DEFAULT_EFFORT=high

Disable Adaptive Thinking

Since February 2026 Claude decides per‑round how much compute to allocate; for tasks judged "simple" it skips the thinking step, leading to downstream bugs. Force a fixed reasoning budget for every round with:

export CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1

Default Permissions Mode

By default each tool call requires manual confirmation, which can flood the UI (e.g., 47 prompts in one morning). Edit settings.json to change the default mode to acceptEdits for trusted repositories or plan for unfamiliar codebases:

{ "permissions": { "defaultMode": "acceptEdits" } }

Allow and Deny Rules

Without explicit rules Claude pauses on basic commands like git status and silently reads .env and .ssh directories. Define clear boundaries in settings.json:

{
  "permissions": {
    "allow": ["Read", "Glob", "Edit", "Bash(git status)", "Bash(npm run *)"],
    "deny": ["Read(**/.env*)", "Read(**/.ssh/**)", "Bash(sudo *)"]
  }
}

Model Switching in Session

Using /model opus for simple regex questions wastes money because Opus costs five times more than Sonnet. Route 80% of everyday coding to /model sonnet, reserve /model opus for complex refactoring, and use /model haiku for quick formatting.

Directed Compression

When the context window fills, a plain /compact call summarises the session and discards crucial architectural decisions. Specify the compression direction to preserve everything you need:

/compact preserve all architecture decisions, file paths mentioned, and error messages

Persist Project Memory

To remind Claude that a project uses pnpm instead of npm, add a memory entry that is loaded automatically at session start:

/memory add "this project uses pnpm, not npm"

Control MCP Token Inflation

The Model Context Protocol (MCP) server can consume over 18,000 tokens per round per connection; five idle servers can burn 90,000 tokens before any prompt is issued. Periodically run /mcp to disconnect unused servers.

Auto‑format After Tool Use

Claude’s generated code often requires a manual Prettier run. Add a PostToolUse hook in settings.json so that any .ts file written by the Agent is immediately formatted:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write(*.ts)",
        "hooks": [{ "type": "command", "command": "npx prettier --write $file" }]
      }
    ]
  }
}

Pre‑tool Log Filtering

Feeding a 10,000‑line server log directly would overflow the context. Use a PreToolUse hook to grep only the most relevant lines before Claude reads the file:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash(cat *log*)",
        "hooks": [{ "type": "command", "command": "grep -n 'ERROR\|WARN' $file | head -50" }]
      }
    ]
  }
}

In summary, Claude Code’s usefulness hinges on these environment variables and the JSON configuration file; spending about a minute to apply them yields far greater gains than swapping models.