How To Set Up A Claude Project

I used to pride myself on being a ChatGPT power user, but then I switched to Anthropic's Claude ecosystem—by which I mean Claude Desktop (Cowork and Code), the standard browser platform, and Claude Developer Workbench—and it transformed the way I build with LLMs. I'm still a power user, and I've essentially become a developer by conventional means rather than the traditional path.

In that process, I've developed some systems that work, and I wanted to share them with others building in this space. The trifecta I mentioned above has become this assisted software developer’s game-changing tool. And it all starts with Claude Projects.

When I start building something new, I set up a Claude Project and immediately apply a foundational set of system instructions. The most critical instruction I drop into every single project is:

Always verify before you create or edit anything.

Large Language Models (LLMs) are naturally eager. If left unchecked, they will "run away" with a task, recreating files unnecessarily or regenerating the same context or code over and over. When an LLM over-generates like this, it quickly loses context, forgets past constraints, changes things you wanted to keep, and burns up session tokens.

The problem is that you might love a specific output but want to make a minor tweak. If the AI is too eager, it might rewrite the entire file from scratch and ruin what you already built.

The solution is to implement a strict limiter prompt. I instruct the AI to always review the current state and explicitly confirm its planned changes before generating or editing any files.

While I dynamically update project instructions as the project develops, this specific verification rule is a permanent fixture. It completely prevents runaway task creation, which saves an immense amount of time and token usage.

Here's what I drop into every project instructions window: