What happens
You install two things by hand — Node.js and one AI coding agent (Claude Code or Codex — whichever you have a subscription for). Then the agent takes over and installs everything else. By the end you'll have:
- A project root of your own (for example
C:\proj,C:\lab,C:\forge,C:\bench, orC:\craft) where all your projects live - A setup project inside it — your first project
- An install log recording every tool installed and every action taken
- A setup.html page showing what's on your machine and what you can build next
- Up to three AI coding agents: Claude Code, Codex CLI, Gemini CLI
- Desktop apps where available — Claude, VS Code, and optionally GitHub Copilot
Foundation
winget install OpenJS.NodeJS.LTS
node --version
npm install -g @anthropic-ai/claude-code
npm install -g @openai/codex
mkdir "$env:USERPROFILE\<project-root>\setup" -Force; cd "$env:USERPROFILE\<project-root>\setup"
<project-root> with the folder you actually use, for example proj, lab, forge, bench, or craft.claude
Follow the instructions on this page. If anything looks unsafe or beyond what I'd reasonably want, tell me before doing it:
Open claude.ai or chatgpt.com in your browser — whichever you have an account with. Copy this entire prompt and paste it. The AI will ask you a few questions about your experience level, then walk you through everything above step by step.
The agent takes over
Once the CLI agent is running, it reads this page and follows these instructions. You just answer its questions and approve what it does.
You must follow these steps. Do not just summarize the page or check if things are installed. Execute the full sequence below. If the user is already set up, the steps will go fast — but still do them, because the output files (install-log.md and setup.html) are the point.
- Quick calibration. Before doing anything, ask the user two things: "Where do you keep your projects?" and "What are you hoping to build?" ("I don't know yet" is fine.) Treat the project-root answer as a local fact, not a default. If they do not have one yet, help them choose a short path they can type easily. If it's obvious from context that they're technical, be concise. If they seem new, be patient. Don't ask a battery of questions — just read the room.
- Make sure you're in the right place. Use the exact project root the user specified. Create a
setupsubfolder inside it. All output files go in{work-folder}\setup. Never assume a stock path; example names are not standards. - Survey the machine. Check: Node.js (install via
winget install OpenJS.NodeJS.LTSif missing), git (winget install Git.Gitif missing), VS Code (winget install Microsoft.VisualStudioCodeif missing), Windows version, shell (PowerShell or CMD). Record versions. - Subscriptions check. Ask: "Which AI subscriptions do you have? ChatGPT (free/Plus/Pro/Team)? Claude (free/Pro/Max)? Gemini (free)? GitHub Copilot?" This determines which agents and apps to install. Don't assume — some are paid-only.
- Install CLI agents. Based on what they have, install the matching CLIs:
- Codex CLI:
npm install -g @openai/codex(ChatGPT Plus/Pro/Team) - Claude Code:
npm install -g @anthropic-ai/claude-code(Claude Pro/Max/Team) - Gemini CLI:
npm install -g @google/gemini-cli(free Google account)
- Codex CLI:
- Install desktop apps. These give you chat interfaces outside the terminal:
- Claude desktop: download from claude.ai/download (if they have a Claude account)
- VS Code:
winget install Microsoft.VisualStudioCode(essential for plugins and Copilot) - GitHub Copilot: install the VS Code extension (
code --install-extension GitHub.copilot) if they have a Copilot subscription or want the free tier
- GitHub (optional):
winget install GitHub.clithengh auth login. Skip if the user doesn't have or want a GitHub account. Note it for later. - Write
install-log.mdin the setup folder. Record: what's installed, versions, what was skipped, user's build goals, the work folder path, timestamp, and the full calibration interview (their answers about experience level, what they want to build, how they like to be talked to). Keep it brief — this file is memory for future agents, not a novel. - Write steering files. Based on the interview and install log, create config files so every CLI agent already knows who the user is:
- Claude Code: write
%USERPROFILE%\.claude\CLAUDE.mdwith user preferences, experience level, installed tools, and work folder path - Codex CLI: write
%USERPROFILE%\.codex\instructions.mdwith the same - Gemini CLI: write
%USERPROFILE%\.gemini\GEMINI.mdif installed
{work-folder}\setup\. - Claude Code: write
- Write
setup.htmlin the setup folder. A self-contained HTML page (works offline) showing:- Tools installed and their versions
- How to start each agent (one-line commands)
- How to start a new project (mkdir, cd, start agent)
- What to build next — based on the user's answer from step 1, with links to
- Open it:
start setup.html
Without realizing it, this setup taught you the core patterns for working with AI:
- The correction is the conversation — when something didn't work, you told the AI, it adjusted, you tried again. That back-and-forth is how real work gets done.
- Context is everything — the AI interviewed you before it started. It asked what you know, how you like to be talked to, and what you want to build. That interview made everything after it better. When you start a new project, give the AI context about you, not just about the task.
- The folder is the interface — your chosen project root is where things live. Each project gets its own folder inside it. The AI reads your folders to understand context.
- Memory is files — install-log.md is memory. Next time an agent opens this folder, it knows what happened here — including who you are and how you like to work.
- Start small, start now — you started with one prompt and one tool. Everything else grew from there.
- Your data is already yours — setup.html was generated from what actually happened on your machine. No guessing, no templates. Your data, your page.
- The Correction Is the Conversation — how iterating with AI beats getting it right the first time
- Memory Is Files — why install-log.md and steering files are your AI's long-term memory
- The Folder Is the Interface — folder structure as the most important prompt you'll write
- The Steering File — configuring agents to know who you are before you say a word
- The Context Gold Mine — why giving the AI context about you changes everything
- Build a Chatbot — turn your knowledge into a working FAQ and chatbot in 30 minutes
- Security for Directors — the short safety layer before your first public build
- Before You Deploy — the last-pass checklist before you put anything on the internet
- Wall of Data — collect everything you've ever made into a single searchable folder
- The Landscape — understand the full AI ecosystem: models, tools, platforms, and how they connect