Tips on Claude

Things on this page are fragmentary and immature notes/thoughts of the author. Please read with your own judgement!

Claude CLI Installation¶

curl -fsSL https://claude.ai/install.sh | bash

Works simiarly to Gemini CLI.

Claude Desktop¶

1. Runs in a sandboxed VM and cannot access local files (which limits its power).

Claude Plugins¶

1. Claude Plugins (similar concept to Gemini extensions) .

General Tips¶

Claude Code cheatsheet

1. Use ctrl + g to edit the prompt using an external editor ($EDITOR).

2. The @ symbol is useful in a few ways.

• Include the content of another file in context files (e.g., CLAUDE.md) so that you don’t have to duplicate the content at multiple places.

• Type in @ within a live claude session brings up a dropdown list for you to select files. This is handy, which avoid typing long file names or copying/pasting file names. As a more flexible alternative, you can press ctrl + g to edit the prompt using Neovim (assume you have the environment variable $EDITOR configured to nvim), which allows you to fully leverage the power of Neovim. For example, you can read in file names under the current directory using :r !ls.

1. Use /review and /ultra-review to review code changes.

2. Enable voice mode using the command /voice. See discussions in claude cli voice input .

3. Generate a long-live token using claude setup-token.

Agent Ochestration (Multi-model Mode)¶

Claude Code Opusplan Model Explained

1. Claude Code has a opusplan model setting which uses Opus for planning and Sonnet for execution. See Claude Code Opusplan Model Explained for details. You cannot see opusplan in the interactive /model menu of the Claude Code CLI because of a known discoverability bug: the interface’s display list omits it, even though the underlying parser fully supports it. To use it, bypass the visual picker menu and type /model opusplan directly into the terminal prompt.

2. The command /batch allows you to research and plan a large-scale change, then execute in in parallel across 5-30 isolated worktree agents that each open a PR.

Permissions¶

1. If you are already inside an active Claude Code CLI session, you can cycle through available permission modes by pressing shift+tab.

   • default: Ask for most actions

   • acceptEdits: Auto-approve file edits and common filesystem operations

   • plan: Read-only planning, no changes

   • auto: Auto-approve actions with safety checks

   There are actually 2 another permission modes dontAsk and bypassPermissions.

   • dontAsk automatically denies every tool call that would otherwise prompt. The only things that run are:

     • Actions matching your permissions.allow rules

     • Read-only Bash commands

     Even explicit ask rules get denied (rather than pausing for you), and writes to protected paths (.git, .claude, etc.) are denied. The result is a fully non-interactive session where Claude can only do what you’ve pre-approved in advance. The user case for dontAsk is locked-down CI pipelines or restricted environments where you whitelist exactly what’s allowed. The dontAsk permission mode never appears in the shift+tab cycle, instead, it can only be turned on by starting Claude CLI with claude --permission-mode dontAsk.

   • bypassPermissions allows everything. It is sometimes called the YOLO (you only look once) permission mode in other agentic tools (e.g., Antigravity CLI). By default, you don’t see bypassPermissions when pressing shift+tab to cycle through permission modes. You have to start Claude CLI with claude --allow-dangerously-skip-permissions to make it show up when cycling through permission modes. You can also directly directly get into the bypassPermissions mode by starting Claude with claude --dangerously-skip-permissions.

     You can start Claude with a specific permission mode toggled on using the option --permission-mode.

2. The "auto"permission mode is generally the recommended setting for non-security-critical tasks. It strikes a good balance between productivity and safety. The “auto” permission mode is a good fit for

   • Refactoring code

   • Writing tests

   • Updating documentation

   • Creating scripts

   • Data analysis on non-sensitive data

   • Local development in personal projects

   • Prototyping and experimentation

   You might prefer a more restrictive mode when:

   • Working with production infrastructure

   • Managing secrets, credentials, or sensitive customer data

   • Running destructive operations (database migrations, large file deletions, etc.)

   • Operating in regulated environments

   And you might use bypassPermissions only when:

   • You’re in a disposable environment (Docker container, VM, sandbox)

   • You fully trust the workflow

   • You want maximum automation and accept the risks

3. The command /fewer-permission-prompts scans your transcripts for common read-only Bash and MCP tool calls, then add a prioritized allowlist to project (not global) .claude/settings.json to reduce permission prompts.

4. You can set the default permission mode in ~/.claude/settings.json . For example,

   "permissions": {
       "defaultMode": "auto"
   }

GitHub Integrations¶

The simplest way to setup GitHub integration is to run the following command.

claude /install-github-app

Notice that there are issues in the generated workflow claude-code-review.yml. Please see discussions in Claude doesn’t generate any output from PR review for root causes and solutions.

• claude-code-action/examples

Anthropic / Claude APIs¶

See Claude API Docs .

References¶

• Claude Code Docs

• Claude Code Docs - Permission Modes

• Stop Fighting Claude Code’s Permission Prompts — Here’s How the System Actually Works

• Claude API Docs