Claude Starter Kit

Three pastes and you have a Claude that knows who you are, how to think about your work, and how to handle complicated projects without losing the plot halfway through.

Built by Josh Bowman after months of trial, error, and a lot of mass-deleting fancy folder structures. This is the version I actually use, stripped down to what works.

What this is

Three things, in order:

1. Personal Preferences. You tell Claude who you are once, in your Claude account settings. It then talks to you like a person who knows you, not a generic AI assistant.
2. A workspace structure. One flat instruction layer: a single CLAUDE.md at the root, a single MEMORY.md, no scattered config files in every subfolder. Content folders inside Projects/, Areas/, Resources/ can nest as deep as your work actually needs. The trap is building a parallel instruction hierarchy. The rule is one place where the AI gets its rules from.
3. The RPI workflow. Research, Plan, Implement (and Validate if the stakes are high). You stop telling Claude to "just go build it" and start handing it a researched, reviewed plan. Output quality goes up. Rework drops.

That's the whole system. If you do research-heavy work and need a real knowledge base, there's a fourth thing about adding Obsidian or Notion. Most people don't need it. See 04-when-you-need-a-vault.md before you build something fancy.

Quick Start (about 10 minutes)

Paste these in order. One at a time. Don't multi-thread.

1. 01-personal-preferences.md - Open a fresh Claude chat. Paste this. Claude interviews you and generates your Personal Preferences block. Paste the output into Claude Settings > Personal Preferences. Save.

2. 02-workspace-claude-md.md - Create one folder on your computer (or in Google Drive, or wherever). Drop a copy of the CLAUDE.md template into the root. Edit the routing table to match your folders. That folder is now your AI workspace.

3. 03-rpi-prompts/ - When you have something bigger than a quick task, run the prompts in order: Research, Plan, Implement. Validate is optional, see the doc for when to add it.

Done. You now have what most people are trying to build with 15 plugins and a Notion database.

Who this is for

You're a knowledge worker (PM, analyst, marketer, researcher, founder, builder) who wants Claude to handle real work, not just answer questions. You've felt the pain of:

• Claude forgetting context halfway through a project.
• Repeating yourself across chats.
• "Build me X" responses that need 6 rounds of fixes.
• Building elaborate folder hierarchies that you never look at again.

If that's you, paste 1, 2, and 3. You'll feel the difference inside an hour.

What's in this repo

claude-starter-kit/
  README.md                       <- You are here
  01-personal-preferences.md      <- Paste 1
  02-workspace-claude-md.md       <- Paste 2
  03-rpi-prompts/
    01-research.md                <- Paste 3 (when you have a real project)
    02-plan.md
    03-implement.md
    04-validate.md                <- Optional 4th step, graduate to this
  04-when-you-need-a-vault.md     <- Read this before adding Obsidian or Notion

A note on where this came from

The RPI workflow is borrowed from HumanLayer's Claude Code commands, which are the gold standard for software engineers using Claude Code. I'm a product manager, not an engineer, so I adapted them for Cowork and for work that isn't all code. If you're a developer using Claude Code, go pull HumanLayer's full versions, they have machinery (Linear integration, GitHub permalinks, codebase agents) that I stripped out for general use.

The PARA folder structure is from Tiago Forte's Building a Second Brain. The "flat over deep" research backing it is in Karl Voit's Don't Do Complex Folder Hierarchies and NN/G's hierarchy research. I didn't invent any of this. I put it in an order that works for me and stripped out what I don't use.

Make it yours

This kit is opinionated, not rigid. The opinions are mine, the work is yours. Things to do once you've got the basics in:

• Edit the knowledge routing table in your CLAUDE.md to match the folders you actually use.
• Add a "Quick Context" section so Claude knows what you do, who you work with, what tools you use.
• Start MEMORY.md empty. Let it fill up when you tell Claude to remember something.
• Resist the urge to scatter instruction files through your workspace. Your content can nest however the work demands. The AI only needs one CLAUDE.md and one MEMORY.md at the root. Once you start adding configs in every subfolder, you're back in Architecture Cosplay.

Updates

• 2026-05-10: Initial release. Includes Validate as an optional fourth RPI step.
• 2026-05-11 (v1.1): Clarified the flat/deep rule. "Flat" applies to the instruction layer (one CLAUDE.md, one MEMORY.md, no scattered configs). Content folders can nest as deep as your work demands. Several friends caught the original phrasing on its first read and asked "why is it making hierarchies"; this update closes that gap.

---

If you're seeing this and want to compare notes, find me on LinkedIn or at joshbowman.dev. Architecture Cosplay is real and it will get you too.