The 3 Things to Set Up Before Any Long Claude Project

Most people start a long Claude project with one folder, no backups, and no rules file. Three weeks in, the AI is wandering, the laptop runs out of disk, and a single corrupted save eats a month of work. None of that is dramatic when it happens. It’s just slow attrition until the project quietly dies.

Three things, set up in one afternoon, prevent almost all of it.

1. A backup story before you write a single file

You don’t need an enterprise backup plan. You need more than one place. The combination that has never failed me: Time Machine for the local snapshot, Google Drive or iCloud for the everyday sync, GitHub for any folder that’s code or text, and Syncthing if you have a second machine.

Why all four? Each one fails differently. Time Machine catches accidental deletes. Drive catches drive failure. GitHub catches your-laptop-was-stolen. Syncthing catches Drive-decided-to-eat-your-files. The full version of this argument lives in your AI workflow is one spilled wine away from gone.

2. A house-rules file the AI re-reads every turn

Drop a file called CLAUDE.md in the root of your project. Anthropic’s official memory documentation describes how the file is loaded automatically at the start of every conversation, so the AI never “forgets” the rules you set.

Inside the file, put the things you’d otherwise have to repeat: the project’s purpose in one sentence, the voice rules, what files are off-limits, what tools to prefer. The full breakdown is in the CLAUDE.md pattern. Without that file, the AI’s tone will drift the moment your earlier instructions scroll out of the context window — the failure mode covered here.

3. A folder structure you don’t have to think about

The structure I keep coming back to:

project-name/
  CLAUDE.md          ← the house rules
  README.md          ← what this project is, in five lines
  source/            ← the actual working files
  drafts/            ← AI-generated drafts before they’re approved
  reference/         ← anything you paste in for the AI to read
  archive/           ← finished things you don’t want to delete
  .gitignore         ← keeps secrets and noise out of GitHub

The split between source/ and drafts/ is the load-bearing one. The AI writes into drafts. You move things into source. That’s your approval gate. Without it, the AI ends up overwriting your real work because there’s nowhere else for it to put output.

Do all three in one sitting

Set up the backup story (~20 minutes). Write the CLAUDE.md, even if it’s 10 lines (~10 minutes). Make the folders (~2 minutes). You’re done. The next time you start a project from scratch, copy this whole structure as a template.

The next layer up — how AI agents actually remember the things you put in CLAUDE.md, given that the model itself has zero persistent memory — is in how AI agents remember things when the AI itself has no memory.

You May Also Like

• Claude Projects: Organize Your AI Workspace
• Claude Skills Explained: Build Custom AI Workflows
• What is CLAUDE.md?
• Claude Usage Limits: How They Work and How to Stretch Them
• AI Glossary: 100+ Terms Every Beginner Needs to Know