Part 2

Setting up the Project

Install Claude Code, build a memory system, and create two skills that will drive "sessions."

Prereqs

A terminal (and be comfortable with basic terminal commands), an IDE (VSCode works), and Node so you have npm / npx. From there you can install Claude Code:

terminal

curl -fsSL https://claude.ai/install.sh | bash

Docs here (stop after step 2). Also install the GitHub CLI and log in. Claude uses it later to push branches and open PRs.

Then cd into your project folder and run claude.

Walkthrough

In this video we build the memory system from scratch. We draft the init prompt in regular Claude, paste it into Claude Code, and end up with a working .memory/ folder.

0:32 → why a memory system matters
1:40 → draft the initialization prompt
2:43 → the folder structure Claude generates
3:12 → Ctrl+G opens your input in your editor
4:27 → --allow-dangerously-skip-permissions + Shift+Tab for bypass mode
5:40 → /resume and renaming sessions

The memory system we're setting up has one job. It makes sure every new session knows (1) what we're building, (2) the current state of the build, and (3) decisions we've already made. Without it, we'd be re-explaining the project every time the context window resets.

A note about leading dots (.memory, .CLAUDE.md): on Mac and Linux, files or folders starting with . are hidden by default. ls and Finder skip them unless you ask. That's why configs live in .memory/, .claude/, .git/. Run ls -a to see them.

The Init Prompt

Skip writing your own. Paste INIT.md into Claude Code and it builds the memory system for you. The prompt does three things. It creates the .memory/ folder with a specific set of tracking files, fills STATUS.md with your stack (HTML/CSS, GitHub Pages, no build step), and stops there. It's essentially a setup script written in English. You'll end up with:

.memory/PLAN.md → phases and "done" criteria
.memory/STATUS.md → current phase
.memory/DECISIONS.md → design choices, append-only
.memory/LAST_SESSION.md → what shipped last time
.claude/ → skills and project config

Skip Permission Prompts

To make the bypass flag from the video (4:27) stick, alias it:

terminal (zsh)

echo "\nalias claude='claude --allow-dangerously-skip-permissions'" >> ~/.zshrc && source ~/.zshrc

The flag is called "dangerously" because Claude can run any shell command without asking, including deleting files, installing packages, or pushing to git. It's fine for this tutorial because you're working inside a fresh project folder and there's no production system to break. Anything Claude does locally can be undone with git. Don't alias this flag in a directory you care about (your home folder, a repo with uncommitted work, anything touching real credentials).

Skills

Skills are reusable prompts you trigger with a slash command. Rather than typing out a long prompt, you can alias it to a skill, so when you run that skill Claude reads that entire prompt. This is how we can start to build out complex workflows in Claude Code that we can run quickly. The actual text for skills live in .claude/skills/<name>/SKILL.md. Set up two:

s-sesh → opens a session. Reads memory, prints a briefing, takes your goal as an argument. SKILL.md
c-sesh → closes a session. Logs what shipped, updates STATUS.md and LAST_SESSION.md. SKILL.md

s-sesh gets you and Claude on the same page. c-sesh hands off to the next session.

CLAUDE.md

CLAUDE.md is a meta-prompt appended to every message in the project, so keep it short. Its job is to remind Claude of the project, the goal, and the memory layout. Drop this one at the project root. It describes the project, structure, rules / constraints we have, how we work with Claude, how we manage memory files, and any conventions we want Claude to abide by.

Continue on to Part 3.