How I use Claude Code

My top 10 tips

I’ve spent the last year using Claude Code for most of my work — data analysis, chart-making, research, writing drafts. Here are the top 10 things I have learned about using Claude Code effectively.

1. Write a CLAUDE.md file

Claude Code reads a file called CLAUDE.md at the start of every session. It’s your instructions file. Mine is about 200 lines — API keys, file paths, color palette, chart formatting rules, writing tone, and behavioral guardrails. Almost every line exists because Claude got something wrong and I wanted it fixed permanently.

The behavioral rules matter most. “Always use D3 for charts, never matplotlib.” “Never fabricate statistics.” “Surface anything that looks off — low match rates, dropped rows, implausible results.” Claude will silently work around problems if you let it, so I require it to list every judgment call it made.

2. Build custom slash commands

Slash commands are saved prompts that expand into full multi-step instructions. If you have a workflow you repeat, encode it once and never think about the steps again.

I have a /new-project command that creates a project folder, writes a starter script with my boilerplate, and creates a linked Notion research page. A /write-article command that gathers all the data work I’ve done in a session, surveys what other outlets have published on the topic, reads my previous articles to calibrate tone, and produces a full article draft for me to edit. A /deep-search command that tells Claude Code to do extensive, multi-round web research — reading sources, following threads, synthesizing across articles — instead of the quick single-query search it does by default.