Advanced Track

If you’re comfortable with the terminal

For participants who already use Cursor, GitHub Copilot, or have shipped projects with Git. Skip the handholding and build something real.

Goal

By the end of the 40-minute hands-on: a functioning mini-project with a CLAUDE.md, a Git repo, a multi-step analysis, and (optionally) pushed to GitHub.

Starting Point

Either use the workshop demo data, or bring your own. If you brought a non-confidential dataset, copy it into your working folder now.

Step 1 — Scaffold a Project

mkdir ~/my-ai-day-project
cd ~/my-ai-day-project
claude

Step 2 — Write a CLAUDE.md Together

Don’t write it from scratch. Let Claude draft it:

“I want to analyze [your research question] using [your data]. I code in R with the tidyverse and fixest. My conventions are [your conventions]. Help me write a CLAUDE.md for this project.”

Review what it writes. Push back. Add your own anti-patterns, your preferred packages, your plotting theme. The best CLAUDE.md files are opinionated.

Step 3 — Plan Before You Code

Van Horn’s rule: “Unless it’s a one-line change, there’s always a plan.md first.”

“Read the CLAUDE.md. Propose a 4-step analysis pipeline for this project. Write it to plan.md. Don’t write any code yet.”

Review the plan. Edit it by hand if needed. This is the checkpoint that survives everything.

Step 4 — Execute the Plan

“Execute step 1 of plan.md. Show me the output before moving on.”

Iterate. Claude will make mistakes. Correct them in plain English. Don’t switch to editing code manually — steer the conversation instead.

Step 5 — Git + GitHub

“Initialize Git. Create a sensible .gitignore for this project. Commit everything so far with a meaningful message.”

“Now help me push this to a new public (or private) GitHub repository. Set the remote and push.”

Step 6 — Multi-Session Workflow

Exit with /exit. Start a fresh session: claude.

“Read CLAUDE.md and plan.md. What has been done? What’s next?”

This is how real workflows actually run. Short focused sessions. Persistent state in files.

Going Even Further

  • Subagents in parallel: Ask Claude to spawn two subagents working on different steps
  • Plan a full paper section: Have Claude read your existing draft and propose a revision as a plan file
  • Replicate someone else’s result: Grab a public paper + data, ask Claude to reproduce their main figure
  • Write your own skill: If you’ve used Claude Code skills, draft one for a recurring task in your workflow

Notes from the Trenches

  • When Claude goes off the rails: Don’t fight the conversation. Type /exit, start fresh, point at your files. State lives in the files.
  • When context fills up: Use /compact. Tell it what to keep: “/compact keep the regression results and variable names.”
  • When the code works but looks wrong: Read it. Don’t trust by default. Your job is to verify, not just accept.
  • When a command is destructive: Read the permission prompt before saying yes.

See Also