90 Days using Claude Code

Ninety days in, AI-assisted development has changed how I scope and ship work. Generalizing, I manage a 80/20 flip: before I was 80% of time manual coding and now barely 20%.

We built IDEs to get away from the terminal. They gave us autocomplete, refactors, visual debugging, and a sense that “modern development” means clicking around in a rich interface instead of typing commands like it’s 1999. That’s why the current wave of AI-powered CLI tools feels a bit paradoxical: after decades of tooling designed to pull us out of the command line, we’re voluntarily going back in

• Context degrades at 40% capacity, not 100% - Most don’t realize this
• CLAUDE.md files: 2.5k tokens good, 15k tokens = Claude ignores half

Common Failure Patterns:

1. Hallucination without verification
2. Infinite loops on same failed approach
3. No definition of done (perfectionism loops)
4. Rewrites code explicitly told not to touch
5. Context window full of failed attempts

1. Opening (the terminal paradox)
2. Week 1: The False Start (Infinite loops on same failed approach, context pollution/rot: Context window full of failed attempts)
3. The Mental Shift (from “instead of me” to “with me”)
4. What Actually Works (Skills, Commands, Scripts)
5. My main skills: statistical-analysis, fklearn, databricks,
6. 90 Days Later (synthesis and honest assessment)