My first interaction with Claude Code:
me: check out this git repo
me: check out this git repo
me: can you check out this git repo?
Three times. I typed the same thing three times because I didn't trust that it heard me. I was talking to a development environment like it was a golden retriever. "Sit. SIT. Can you sit?"
This is what Level 1 looks like. Everyone starts here. If you're reading this and thinking "I would never," you absolutely would, and you probably did.
The Permission Problem
My early prompts had a specific energy:
"Can we use AWS CDK to manage how this is deployed?"
Can we. I was asking a tool for permission. Like I was pitching a risky idea at sprint planning and needed buy-in from a coworker who has opinions about infrastructure choices. Claude Code does not have opinions. It does not need to approve your technology selection. It will use CDK or Terraform or a shell script that deploys via carrier pigeon if you tell it to.
But years of chatbot muscle memory kicked in:
- Asking questions instead of giving instructions. "Can we..." instead of "Use CDK."
- Repeating yourself because you're sure the context evaporated between messages.
- Explaining your reasoning like you're defending a design document. Claude does not care why. It cares what.
Here's what the shift looks like:
Before:
"Can we use AWS CDK to manage how this is deployed? I was thinking we could set up a basic stack with S3 and maybe CloudFront if that makes sense for this use case."
After:
"Set up a CDK stack. S3 bucket with CloudFront distribution. TypeScript."
Same result. 80% fewer words. Zero emotional labor.
The Trust Document
Claude Code has a .claude/settings.local.json file that controls what it's allowed to do. My permissions history is basically a diary of trust issues:
Week 1: One permission. WebSearch. I was using the most capable coding tool on the market to google things.
Week 3: Two permissions. Create directories and run exactly one prompt file. I was planning Thanksgiving side dishes with an AI coding assistant. The technology-to-task ratio was approximately 10,000:1.
Month 2: 150+ explicitly allowed commands, including cdk:destroy. Each one earned by clicking "allow" on the confirmation prompt and getting sick of it. The settings file is a document of trust built one irritated click at a time.
The CLAUDE.md Gap
Every Claude Code project has a CLAUDE.md file, the README that Claude actually reads. Mine said:
## Project Status
This is a new repository. Update this file as the project develops.
"Update this file as the project develops." Narrator: it was never updated. Not here, anyway. It took me until Post 8 to write a real one.
The Takeaway
Level 1 is treating Claude Code like a chatbot. You ask permission. You repeat yourself. You explain your reasoning to software that doesn't have feelings. Everyone does this. The interesting stuff starts at Level 2, when you stop copying Claude's output and start reviewing it. When you stop asking "can you write this function" and start saying "build the auth system."
But you can't skip Level 1. You have to ask the golden retriever to sit three times before you learn that it was already sitting.