AI Coding Agent Tips

🖥️

Terminal Setup

Terminal theme and input optimization

Switch themes with /config, enable Shift+Enter line breaks with /terminal-setup. Vim mode also available with /vim.

⚡

Effort Levels

Token usage vs intelligence level control

Select Low/Medium/High effort levels with /model command. Low is fast and cheap, High is slower but smarter.

NEW

🚀

Permanently Set Effort Level to Max

Force max via env var — bypassing the settings.json bug

Since Claude Code v2.1.68 (2026-03-04), the default effort level changed to medium, potentially lowering answer quality. Adding `export CLAUDE_CODE_EFFORT_LEVEL=max` to `.zshrc` applies it permanently across all sessions. settings.json has a max persistence bug, and CLI flags need to be typed each time — env var is the cleanest solution.

🔌

Plugins, MCP, Skills

MCP server, skills, and agent extensions

Install LSP, MCP, skills, and agents from the plugin marketplace with /plugin. You can also create a company-private marketplace.

🤖

Custom Agents

Create custom agents in .claude/agents

Add .md files to .claude/agents/ to create custom agents. You can specify name, color, tools, permissions, and model individually.

🔐

Permission Pre-approval

Accelerate workflow with permission pre-approval

Set per-tool allow/block lists with /permissions. Supports wildcards — Bash(bun run *), Edit(/docs/**), etc.

📦

Sandbox Mode

Enhanced safety with file/network isolation

Activate open-source sandbox runtime with /sandbox. File isolation + network isolation reduces permission prompts.

📊

Custom Status Line

Display custom info below the composer

Display model name, directory, remaining context, cost, etc. right below the composer with /statusline.

⌨️

Keybinding Customization

Freely remap all keybindings

Remap all keybindings in Claude Code with /keybindings. Settings are hot-reloaded for instant effect.

🪝

Hooks

Deterministically intervene in the Claude lifecycle

Route permission requests to Slack via hooks, add nudges at turn end, pre/post-process tool calls, and more.

🔄

Spinner Verbs

Customize loading spinner verbs

Add or replace your own verbs in the default spinner verb list via settings.json. Shareable with your team.

🎨

Output Styles

Configure response tone and format

Set output style in /config. "explanatory" style explains code patterns while working, "learning" style provides coaching.

⚙️

settings.json Management

37 settings + 84 environment variables

Check in settings.json to Git to share team-wide settings. Configurable at codebase/subfolder/personal/enterprise level.

📝

CLAUDE.md Writing Guide — Order Optimization & Best Practices

Stable content on top, frequently changing content at bottom — Prompt caching optimization

CLAUDE.md is the project rulebook auto-loaded every session. Order matters — put stable content (project description, coding rules) on top and frequently changing content (current tasks) at bottom for better prompt cache hit rates.

⚖

Commands vs Skills Complete Comparison

.claude/commands/ vs .claude/skills/ — When to Use Which

Commands are simple single-file prompts, Skills are complex directory-structured workflows. They differ in invocation, structure, and auto-triggering. Side-by-side comparison.

🧩

Context · Memory · Handoff Complete Guide

Three "Memory" Mechanisms in Claude Code — When Each Is Used and How They Differ

Claude Code has three memory systems: Context (current conversation), Memory (cross-session), and Handoff (session transfer). Confusing these three causes "why doesn't it remember?" problems.

🧠

Auto Memory

Automatic context retention across sessions

Claude Code auto-saves important info across sessions to ~/.claude/projects/. Manual saving also possible with "remember this" command.

💬

/btw (By The Way)

Ask side questions without polluting the main conversation

The /btw command lets you ask separate questions on the side without interrupting the current conversation. Similar to opening a separate chat in Cursor, now built into Claude Code.

🤖

Auto Permission Mode

Auto-judge tool usage for long autonomous work sessions

With claude --permission-mode auto, Claude judges tool usage by itself, enabling long autonomous work sessions without human approval.

🎙️

/voice — Voice Input

Give instructions to Claude Code by voice

Start voice input mode with /voice command. Give instructions in natural language without a keyboard for hands-free coding.

🗜️

/compact — Context Compression

Compress conversation context to save tokens

Use /compact to summarize and compress the current conversation, freeing up context window. Continue working without performance degradation even in long sessions.

📱

Remote Control — Claude Code on Mobile

Access from smartphone/browser with claude remote-control

Use claude remote-control to continue your local session from your smartphone (Claude app) or browser (claude.ai/code). Connect by simply scanning a QR code without SSH setup.

🔄

--resume / --continue — Resume Sessions

Continue previous conversations, move sessions between claude.ai and Claude Code

Instantly resume the most recent session with --continue, or select from past sessions with --resume. Move sessions between claude.ai web and Claude Code.

🚫

.claudeignore — Block Unnecessary File Reading

Prevent token waste + protect sensitive files

.claudeignore uses the same syntax as .gitignore to specify files/directories Claude Code should not read. Blocks large files, build outputs, and sensitive files to prevent token waste.

🎓

Claude Code Learning Resources Guide

Official courses, hands-on labs, MCP & Agent skills — Best links for learning

Curated list of Anthropic official learning platform (Skilljar) and GitHub course repos. Step-by-step learning path from Claude Code basics to Agent Skills and MCP.

📨

Channels — Control Claude Code from Telegram/Discord

Send messages to your session from mobile — Claude works on your local filesystem

Channels feature launched March 2025. Send messages from Telegram or Discord to a running Claude Code session — Claude accesses your local files, Git, and MCP to work and replies back. Not cloud — just opening another window to your local session.

📋

.cursorrules File

Pass project rules to Cursor

Record coding style, framework rules, and prohibitions in .cursorrules and Cursor will automatically follow them.

🎵

Composer (Agent Mode)

Multi-file automatic editing

Open Composer with Cmd+I to use agent mode that edits multiple files simultaneously. Agent mode can also execute terminal commands.

↹

Tab Completion

Optimized code auto-completion

Press Tab to auto-complete predicted next code. Supports multi-line prediction and diff-based edit suggestions.

📄

.github/copilot-instructions.md

Pass project rules to Copilot

Define coding rules in .github/copilot-instructions.md and Copilot Chat will automatically reference them.

💬

Chat Agents (@workspace)

Use the entire codebase as context

Search the entire codebase and answer questions with @workspace agent. @terminal and @vscode agents also supported.

🏗️

Copilot Workspace

Issue → Plan → Code automatic workflow

Click "Open in Workspace" from a GitHub Issue to auto-plan → code changes → PR creation.

📊

Project Context File Comparison

CLAUDE.md vs .cursorrules vs copilot-instructions

Each tool has a different file for passing project rules. The core content is the same, so create one well and copy it.

✍️

Effective Prompt Writing

Give clear instructions to AI coding agents

Mention specific file names and function names, clearly describe the desired output format. "Change Z to W in Y function of X file" instead of "fix it".

📐

Context Window Management

Work efficiently within token limits

Exclude unnecessary files from context and start new sessions for large tasks. Compress context with /compact in Claude Code.

🧪

Automate Test Writing with AI

Request tests right after writing code

Request "write tests for this code" right after writing code — generates accurate tests while context is fresh.