Claude Code is all you need

Let’s deep dive into Claude Code.

Table of Contents

Use this to find all env vars that Claude Code supports

$ rg -o "process\.env\.\w+" $(which claude) | sort -u | sed 's/process\.env\.//'
ALACRITTY_LOG
ALIYUN_REGION_ID
ANTHROPIC_API_KEY
ANTHROPIC_AUTH_TOKEN
ANTHROPIC_BASE_URL
ANTHROPIC_BETAS
ANTHROPIC_CUSTOM_HEADERS
ANTHROPIC_MODEL
...

Articles

Architecture and prompt analysis

https://southbridge-research.notion.site/claude-code-an-agentic-cleanroom-analysis

Claude Code decompiles itself

https://ghuntley.com/tradecraft/

Context Engineering

https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html

https://www.dbreunig.com/2025/06/26/how-to-fix-your-context.html

Hook

Official doc: https://docs.anthropic.com/en/docs/claude-code/hooks

Articles:

Debugging with tmux and GDB

Run GDB in a tmux session while Claude Code sends debugging commands. Attach to the session from another terminal to monitor output.

Example

# Start a tmux session with GDB
tmux new-session -d -s debug "gdb ./program"

# Claude Code sends commands to the debugging session
tmux send-keys -t debug "break main" Enter
tmux send-keys -t debug "run" Enter

# Monitor debugging output in another terminal
tmux attach-session -t debug

Voice Input with SuperWhisper

Use SuperWhisper for voice-to-text input with Claude Code. Dictate complex commands and code naturally instead of typing.

Update on 2025-10-01, with Sonnet 4.5

On 2025-09-29, Anthropic released Sonnet 4.5, claiming it to be smarter, faster than Opus 4.1.

At this point, I has been frustrated by Opus 4.1 in my work for about 1 month. The pain point is that Opus 4.1 is

• very slow
• expensive
• don’t always follow the instructions
• lazy at gathering context and jump right into the implementation, in other words, right new code.

First impressions on Sonnet 4.5:

• I tested it in using Anthropic Max subscription. Will test it via Vertex AI during work.
• It’s way faster, thanks to 1. lower inference latency 2. ability to call tools concurrently

One typical scenario is that I keep inputting messages to the queue and I can see it keep acking more messages without stopping modifying code. In previous versions, this manuver will lead to claude missing some of my instructions.

Quota limits

With ccusage and https://claude.ai/settings/usage, I can see the 5-hour limit for Max plan is about $150 worth of tokens. I’m not sure how the weekly limit will kick in.