Teaching Arthur
Claude Code

# Context Document: Claude Code Lesson for Arthur This document provides comprehensive context for an LLM assistant helping with a Claude Code teaching session. Paste this into a conversation to get the assistant up to speed immediately. --- ## The Situation **What:** Aaron Bergman is being paid (as a contractor by Arthur's employer) to teach Arthur Wright how to use Claude Code in a 1–1.5 hour in-person lesson. **When:** February 27, 2026. **Where:** In person, sitting next to each other. Both on MacBooks (macOS). **Why Arthur asked:** He thinks Aaron is exceptionally good at building things and highly leveraging Claude Code despite not writing any code himself. Arthur wants to learn this skill. --- ## About Aaron (the teacher) - Power user of Claude Code. 187 sessions, 772 hours of compute, 77 analyzed sessions. - **Does not write code.** Directs Claude Code to build things for him. This is the key skill he's teaching. - Builds interactive web apps (sandpile sims, Bezier editors, synthesizers, particle life, L-systems, etc.), data analysis pipelines, audio processing tools, web scrapers, and more — all via Claude Code. - Style: states a goal, lets Claude take the first swing, course-corrects when it goes wrong. Rapid iteration. Comfortable interrupting. - Has built an autonomous Claude Code loop with token budget safety caps. - Top friction: Claude choosing wrong initial approaches (28/77 sessions) and buggy first drafts (19/77 sessions). Rarely a misunderstanding problem — Claude gets what Aaron wants, just sometimes picks the wrong technical path. - Uses Python heavily (355 file touches), plus HTML/JS/CSS for web apps (134 HTML file touches). - Uses `uv` as Python package manager, prefers progress bars, prefers scripts that just run with `python script.py`. --- ## About Arthur (the student) - **Arthur Wright**, Head of Community at Effective Altruism DC (since July 2024). - Previously: Graduate Research Fellow at Mercatus Center / George Mason University (MA Economics, 3.93 GPA), Research Volunteer at PolicyEngine. - BA in Philosophy and Religious Studies from Wesleyan University. - **Digital native, smart guy.** Uses AI chatbots (Claude/ChatGPT web) regularly. - **Has dormant programming background:** Python class in high school, AP Computer Science (C++), and Stata for data analysis in grad school. Has IDLE installed (so Python is likely on his machine). But says he's forgotten basically all of it. - **Has seen the "Code" tab in the Claude desktop app**, which has some agentic coding capability now. Worth clarifying early that the terminal-based Claude Code is more powerful and flexible — that's what the lesson covers. - **GitHub scares him** — his words. Keep git/GitHub minimal and pressure-free. - Probable use cases: building event pages for EA DC, data analysis on community/survey data, personal website, cost-effectiveness calculators, simple tools for his work. --- ## About the Lesson **Goal:** Arthur can independently build useful things with Claude Code (web apps, scripts, data analysis) and knows enough to keep learning on his own. **Format:** In-person, both on MacBooks, sitting side by side. Aaron can lean over and help directly on Arthur's machine. **Key principle:** This is NOT a coding lesson. It's a lesson in directing an AI that codes. The skill being taught is: knowing what you want, describing it clearly, and recognizing when the output is right or wrong. --- ## Files in This Project | File | Description | |------|-------------| | `curriculum.md` | Full lesson plan in markdown | | `lesson.html` | Polished infographic-style HTML version of the curriculum (dark theme, color-coded phases, interactive checklist) | | `report.md` | Aaron's Claude Code usage insights report — detailed analysis of how he uses Claude Code | | `context.md` | This file | --- ## Full Curriculum ### Before You Start: Quick Assessment (2 min) Ask Arthur these to calibrate what to skip: - "Have you ever used Terminal on your Mac?" → determines Phase 0 depth - "Do you have VS Code or any code editor installed?" → determines setup scope - "Have you ever used git or GitHub?" → determines whether to cover version control - "What kind of things are you hoping to build?" → shapes which demo/exercise to pick --- ### Phase 0: Setup & Prerequisites (0–25 min, skip what's done) This phase might take 5 minutes if he's set up, or 25 if starting from scratch. Do it together — drive on his machine or coach him through it. #### The Essentials Checklist 1. **Terminal basics** - Check: Can he open Terminal? (Cmd+Space → type "Terminal") - If not: Show Spotlight search, open Terminal, explain it's how you "talk to the computer directly" - Cover: `pwd` (where am I), `ls` (what's here), `cd folder` (go in), `cd ..` (go back), `cd ~` (go home) - Remember: Show that the "path" in Terminal corresponds to folders in Finder. Open Finder side-by-side. 2. **Homebrew** - Check: `brew --version` - Install: `/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"` - Remember: After install, Homebrew prints "Next steps" with PATH commands. Don't skip. 3. **Node.js** - Check: `node --version` (need v18+) - Install: `brew install node` 4. **Claude Code** - Check: `claude --version` - Install: `npm install -g @anthropic-ai/claude-code` - Auth: Needs Claude Pro/Max subscription or API key. First run opens browser login. Make sure he's logged into claude.ai first. 5. **Code editor** - Check: VS Code, Cursor, or similar? - Install: `brew install --cask visual-studio-code` - Remember: Install shell command so `code .` works (Cmd+Shift+P → "Shell Command: Install 'code' command") 6. **Git** (nice to have) - Check: `git --version` (often pre-installed) - Install: `xcode-select --install` - Don't get into git workflows. Just make sure it's installed. 7. **Python + uv** (optional) - Check: `python3 --version`, `uv --version` - Install: `brew install uv` - Only if Arthur's interested in data/scripting. #### Troubleshooting - `npm install -g` permissions error → `sudo npm install -g @anthropic-ai/claude-code` - Terminal font too small → Cmd+Plus - Anything taking >5 min → let it run in background, move on --- ### Phase 1: The Mental Model (5–10 min) The most important part. If Arthur gets this, everything else follows. **What Claude Code is:** "You know how you use Claude in a browser? Claude Code is like that, except Claude can also see your files, edit them, run commands, and browse the web. It's like having a programmer sitting next to you who can actually touch your computer." **The key insight:** "You don't need to know how to code. You need to know what you want. Claude writes the code, you direct it. Your job is to: (1) say what you want, (2) watch what it's doing, (3) tell it when something's wrong, (4) tell it when something's good." **How it's different from web chatbots:** | ChatGPT / Claude web | Claude Code | |---|---| | You copy-paste code snippets | Claude directly creates and edits files | | Conversation only | Can run programs, open browsers, install packages | | Starts from scratch each time | Reads your project, remembers context via CLAUDE.md | | You implement suggestions | Claude implements, you approve | **Permission model:** Claude asks before doing things. Allow once, or allow for session. Start cautious, get more permissive as comfort grows. --- ### Phase 2: Live Demo — Aaron Drives (10 min) Aaron builds something on his own laptop while Arthur watches. Lets Arthur see the full flow without pressure. **Script:** 1. `mkdir ~/demo && cd ~/demo` 2. `claude` 3. Show the interface 4. Prompt: "Build me a single-page web app that visualizes the trolley problem as an interactive decision tree. When you click a choice, it branches and shows the next dilemma. Make it look clean and modern. Use a single index.html file." 5. Narrate as Claude works 6. Open result in browser **Point out:** "I didn't tell it how, just what" / "Watch the tool calls" / "I can interrupt with Escape" / "If something looks off, I'd just say so" --- ### Phase 3: Arthur Drives — Guided First Build (15–20 min) The core of the lesson. Arthur builds something with Aaron coaching. **Setup:** Terminal → `mkdir ~/my-first-project && cd ~/my-first-project` → `claude` **Project ideas:** - EA/Community: Event landing page with RSVP - Economics: Compound interest calculator with chart - Fun: GDP guessing game between countries **Coaching cues:** - Permission prompts → "Go ahead and allow it" - When done → "Let's look at it. `open index.html`" - Something wrong → "Tell Claude in plain English" - Unsure what to type → "Describe what you see vs. what you want" **The iteration loop (do 2-3 rounds):** Look → find issue → tell Claude → see update → repeat. This IS the skill. --- ### Phase 4: Key Skills & Power Features (10–15 min) 1. **CLAUDE.md** — persistent project instructions. Like a sticky note Claude reads every time. Also global `~/.claude/CLAUDE.md`. 2. **Screen layout** — VS Code (left, with integrated terminal running Claude Code) + browser (right). VS Code matters for seeing files change in real-time. 3. **Starting fresh** — `/clear` or Cmd+K when confused or conversation is long. Continue when iterating. 4. **Shortcuts** — Ctrl+` (VS Code terminal), Cmd+P (open file), Cmd+Shift+P (command palette), Escape (interrupt Claude), /clear, /help 5. **Error handling** — Paste errors back to Claude. If circling, say "try a different approach." If stuck, start fresh. 6. **Cost** — Pro/Max subscription includes usage. API costs per message. Don't optimize early. --- ### Phase 5: Independent Build (10–15 min) Arthur builds solo. Aaron nearby for questions but tries not to intervene. **Ideas:** EA event page, cost-effectiveness tool, CSV dashboard, personal site, feedback form. **Watch for:** Can he describe wants? Does he iterate? Does he read tool calls? Does he interrupt when needed? Let him struggle a little — that's where learning happens. --- ### Phase 6: What to Do Next (5 min) 1. Build something every day for a week, even 15 minutes 2. Try data analysis with a CSV he has 3. Try web scraping 4. Set up CLAUDE.md files for projects **The one thing:** "You're not learning to code. You're learning to direct an AI that codes." --- ### Cheat Sheet: Things Easy to Forget - Homebrew PATH commands after install — don't skip - VS Code `code` shell command installation - Arthur logged into claude.ai before first `claude` run - Always `cd` into project folder before running `claude` - Finder ↔ Terminal: drag folder into Terminal for path, `open .` to open Finder from Terminal - macOS security dialogs on first run — click Allow - `open index.html` to view web pages - Explain "localhost" if starting a server - Terminal font size: Cmd+Plus - #1 beginner mistake: running `claude` from home directory instead of project folder - Copy-paste works in macOS Terminal (Cmd+C / Cmd+V) - npm EACCES error → `sudo npm install -g` - Demo your own setup (CLAUDE.md, workflow, screen layout) - Let him struggle — don't intervene too fast --- ## Aaron's Usage Insights (for reference) Aaron's Claude Code usage analysis (187 sessions, 77 analyzed) reveals: **What works well:** - Autonomous web app generation with Playwright verification (15+ apps, high success rate) - Turning messy data exports into polished dashboards and reports - Quick to recognize wrong approaches and redirect **Main friction:** - Wrong initial approach (28 sessions) — biggest time sink - Buggy first drafts (19 sessions) - Local environment/tooling struggles (port conflicts, wrong directories, missing deps) **Key patterns:** - States goals, lets Claude take first swing, iterates - 772 hours of compute with only 705 messages — heavy autonomous usage - Python-heavy (355 file touches) + HTML/JS/CSS for web apps (134 HTML touches) - Satisfaction high despite friction (82/96 at "likely satisfied" or above) **Things Aaron does that Arthur should learn:** - Describe the "what" not the "how" - Interrupt early when Claude goes wrong - Iterate 2-3 times on output rather than accepting first result - Use CLAUDE.md for persistent context - Use plain English — technical vocabulary is unnecessary