Starting to Use Claude Code

Claude Code
Agents
Author

Lawrence Wu

Published

July 18, 2025

I had tried Claude Code briefly when it first came out 4 months ago. However I wasn’t that impressed. Lately though, I’ve seen more and more people turn to Claude Code. The step change was the release of Claude 4 Sonnet and Opus on May 22, 2025. I think something was unlocked where given a powerful enough reasoning model (Claude 4) and access to generic tools, Claude Code is an amazing tool.

Some of the things I read/watched that motivated me to try it again:

Set Up

At work, we use VertexAI. Google has signed an agreement with Anthropic to license Claude models that you can then call via VertexAI. The instructions are easy to follow. In my ~/.zshrc, I just have:

export CLAUDE_CODE_USE_VERTEX=1
export CLOUD_ML_REGION=YOUR_GCP_REGION
export ANTHROPIC_VERTEX_PROJECT_ID=YOUR_GCP_PROJECT_ID
export ANTHROPIC_MODEL='claude-sonnet-4@20250514'

You then authenticate with Google Cloud:

gcloud auth login

Install Claude Code:

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

Then in your project directory or a new directory, start Claude Code:

claude

I eventually started running:

claude --dangerously-skip-permissions

First Experiments

While I was on PTO during the July 4th week, I used Claude Code to build an application I’ve always wanted internally at my company. I had wanted to make a website using FastHTML since I had heard it was much more flexible than Streamlit/Gradio apps. So I finally got around to it thanks to Claude Code. It also figured out you can combine a FastHTML and a FastAPI service because they are both Starlette applications.

The second thing I built was an MCP Server using FastMCP that implemented bearer token authentication.

Both of the above tasks I estimate would’ve taken me 2-3 days to do but it take Claude Code <2 hours with ~5 prompts to implement.

Some best practices I’ve arrived at

  • Right now I just run Claude Code in a terminal. Sometimes I run it in an IDE (VS Code, Windsurf) where the UI is a little different.
  • claude --dangerously-skip-permissions so Claude doesn’t have to ask for permissions.
  • Run the slash command /init to create CLAUDE.md
  • add links to documentation in your prmopts so Claude Code can use it’s fetch tool to learn how to use APIs. Many services/projects have llms.txt that are a good starting point. For example when working with FastHTML, I passed it this.
  • ask Claude to make plans and write those plans in a markdown file in docs/ with a date, e.g. docs/plan_20250715_implement_feature_a.md with ‘ultrathink’ which increases the thinking budget to the max
  • Then you can ask Claude Code to implement the plan
  • Make commits after each successful run of the agent
  • Though today I attended a Claude Code Camp hosted by Every. I learned a ton about aliases, slash commands, how you can trigger subagents in Claude Code and how to think about the 4 step loop of development: planning, delegating, assessing and codifying.

Claude Code Utilities