Hidden Settings To Unlock The True Potential Of Your Claude Code Setup

Highlights

• Setting 'Claude Code New Init' to 1 creates an interactive initialization process that generates custom Claude.md files, hooks, and skills based on workflow answers.

• The 'prompt audit' command uses sub-agents to scan instruction files and remove outdated prompt rules optimized for older model context.

• Setting 'Disable Claude AI Connectors' to true in settings.json prevents background applications like Notion and Slack from eating up session context window space.

• The 'Ask User Question Timeout' setting automatically proceeds with recommended choices after a set duration (e.g., 60 seconds) when an agent is unattended.

• Running the 'model Opus plan' configuration limits the higher-cost Opus model strictly to plan mode while executing code tasks using Sonnet.

Timeline

1. 00:00 - 02:32

   Project Initialization and Migration Commands

   Default initialization overwrites Claude.md with code context that the model can already read directly from the repository. Setting the environment variable triggers an interactive survey about Git workflows and deployment plans before generating files. The import command preserves external setups by converting third-party instructions into native Claude Code skills automatically.

   • Setting Claude Code New Init to 1 prevents the initialization step from dumping automatically derived codebase facts into Claude.md.
   • The import command transfers existing MCP configurations and instruction files directly from other coding tools like Cursor or Codex into Claude Code skills.
2. 02:33 - 04:12

   Context Optimization and Skill Maintenance

   System instructions tailored for legacy models waste space inside current context windows and harm response quality. Sub-agents execute audits to clean up unused rules following model releases. Skill Doctor and sorting shortcuts isolate heavy or rarely called skills so they can be deleted from active memory.

   • Running prompt audit evaluates existing system prompts and strips obsolete instructions that degrade newer model outputs.
   • The Skill Doctor command provides token footprint statistics and execution history for every active skill.
   • Pressing T within the skills list sorts configured skills in ascending order based on token consumption.
3. 04:13 - 06:28

   Permission Rules and Connector Management

   Unnecessary confirmation prompts stall long-running automated tasks. Updating permission settings with auto mode setup and fewer permission prompts allows harmless read commands to execute without user intervention. Disabling background connectors and browser artifacts recovers context space that is otherwise claimed at session start.

   • Auto mode setup updates execution permissions by analyzing historical command approval patterns.
   • The fewer permission prompts command extracts read-only operations from past sessions and appends them to a project allow list.
   • Setting Disable Claude AI Connectors to true stops background integration servers like Notion or Slack from loading into every session.
4. 06:29 - 10:13

   Model Evaluation, Planning Controls, and Output Styles

   Automated evaluation tools allow continuous testing of model performance against fixed criteria. Context retention from long planning sessions can pollute code execution steps, making context clearing after plan approval beneficial. Setting custom plan directories helps teams share design plans inside repository folders.

   • The build eval and hillclimb commands construct test suites and iteratively benchmark AI application changes.
   • Setting show clear context on plan accept restores the option to wipe chat history before execution starts.
   • Configuring plans directory diverts saved plan files into a specified project directory instead of the central .claude folder.
   • The output style command switches model behavior between proactive, concise, explanatory, and learning modes.
5. 10:14 - 13:57

   Session Control, Focus Modes, and Multi-Agent Messaging

   Managing long or concurrent terminal tasks requires granular session isolation and noise reduction. Focus mode cleans the terminal display to minimize scrolling during extended sessions. Inter-agent messaging allows separate instances working on different parts of a project to request status updates directly without manual copy-pasting.

   • Enabling show thinking summaries in settings.json exposes the underlying reasoning chain behind responses.
   • Focus mode eliminates intermediate step output to present only the user prompt, final response, and action summary.
   • The branch, fork, and subtask commands handle context isolation across separate background processes and child sub-agents.
   • Sessions communicate directly with each other using the list agents command and targeted inter-session messaging.
   • Setting Ask User Question Timeout causes unattended sessions to select default options when stalled on non-critical choices.
6. 13:58 - 16:21

   Model Selection Controls and Prompt Keyboard Shortcuts

   Splitting model assignments preserves quota limits on lower-tier accounts by running compute-heavy planning on high-tier models while delegating code generation and child tasks to faster models. Integrated terminal shortcuts prevent lost input when clearing or editing long prompt strings.

   • Running model Opus plan restricts Opus usage exclusively to planning phases while using Sonnet for general execution.
   • Sub-agent models can be globally locked to smaller models using the subagent model force environment variable.
   • Control keyboard shortcuts enable prompt stashing (Ctrl+S), text restoration (Ctrl+Y), and prompt undo (Ctrl+_).
   • The powerup command launches interactive, animated feature demonstrations within the terminal interface.