Anthropic Claude Code CLI: Prompts & Tool Definitions

05 Mar, 2025

Anthropic has recently released Claude Code - an agentic coding tool (interactive CLI)

You can play around it by installing it in your machine

npm install -g @anthropic-ai/claude-code

But beware, It is not free and you get charged based on your token usage (Anthropic API)

This tool is not open source. And you get minified version of the CLI when you install it.

However, I recently came across a tweet by Travis Fischer where he has shared the unminified version (not complete?) which he has extracted from npm.

And the following key learning:

User messages are checked for phrases to activate different levels of thinking mode
- think harder, think intensely, think longer, think really hard, think super hard, think very hard, think about it, think a lot, think hard, think more, megathink (🤣, those words are still there in the codebase)
- higher levels = more tokens
Tech Stack
- React CLI App built using commander and ink
- Zod for validation
- Sentry for error logging
- Has third party inference support
“tengu” seems to be internal Anthropic codename for this project
KISS principle when with respect to file context, embedding, RAG or Abstract Syntax Trees
- Nothing fancy things
- Just plain old - glob search for filenames (which is what even IDE like Cursor does)
- ripgrep for file contents
They support basic planning via the Architect Agent tool and sub-agent invocations via the Task tool (aka dispatch_agent) – bare minimum at the moment
Their documentation list of tools matches the implementation.
They auto-create a CLAUDE.md file to serve as memory for project structure, guidelines, conventions, and common commands.
- When creating this file, they’ll existing prompts like Cursor rules or GitHub Copilot (no Windsurf at the moment)
They check for potential malicious intent (like malware creation) and decide whether to process the request or not.

Happy understanding Claude-Code!