Pair Programming with AI in the Terminal: My Claude Code Experience

Recently I tried Claude Code, a terminal‑based AI programming tool that differs from typical AI assistants by running directly in the shell, understanding full project context, and handling many coding tasks autonomously.

Smooth start with configuration – The first hurdle is API access limits, but Claude Code allows custom base_url and api_key, enabling proxy services such as the popular claude-code-router (CCR) , which converts API formats for easy integration.

Another advantage is that the open‑source Kimi K2 model from Moonshot now supports the Anthropic protocol natively, so switching to K2 only requires an environment‑variable change, unlocking its strong agentic capabilities inside Claude Code.

Effective use of Claude.md – Creating and updating a Claude.md file acts as a project brief for the AI. By documenting the tech stack, directory layout, core logic, and coding conventions, each conversation starts with full context, saving time and reducing token consumption.

Parallel development with subagents – Claude Code’s subagent feature lets multiple Claude instances run in separate terminal windows. One instance can focus on front‑end components while another handles back‑end APIs, markedly improving development speed for loosely coupled modules.

Comparison with other tools – Compared with GeminiCLI‑based tools (e.g., QWEN Code), Claude Code reads code more precisely by using offset‑based selective loading instead of loading the entire repository, which cuts token usage at the cost of some global view. For exhaustive code analysis, GeminiCLI may still be preferable. Against IDE‑centric tools like Cursor, Claude Code offers greater flexibility and lower cost, especially when the subscription includes free token quotas.

Practical tips

Distinguish tasks that can run autonomously from those needing supervision.

Commit code before letting Claude work so you can roll back if the AI drifts.

Provide detailed prompts specifying inputs, outputs, performance expectations, and style.

Ask Claude to generate tests for its code, but be aware of its limits to avoid endless “debug loops”.

Leverage parallel subagents for low‑coupling modules, but synchronize progress regularly to prevent integration conflicts.

Conclusion – Claude Code serves as an always‑ready pair‑programming partner that handles repetitive work, allowing developers to focus on architecture and problem solving. The emerging “Vibe Coding” mindset—describing desired functionality in natural language—becomes feasible when AI tools like Claude Code are deeply integrated with the filesystem, test runners, and Git.

For those interested, start by configuring Claude.md, try simple tasks, and gradually explore the tool’s boundaries.

Resources

Anthropic team’s Claude Code usage experience: https://www.anthropic.com/news/how-anthropic-teams-use-claude-code

ClaudeLog – deep analysis of Claude Code: https://claudelog.com/

Claude Code Router project: https://github.com/musistudio/claude-code-router