The AI Pair Programmer

Building ArchAIc with Claude Code partly to understand how it works and what my teams are doing, how they're using it. What does it do well, and what does it do... not so well.

Can't evaluate AI pair programming without experiencing it hands-on. Can't set policies around AI coding assistants without knowing how they work. Team leaders ask "should we let juniors use AI tools?" and I can't answer from theory. ArchAIc is my data source.

Started about a week ago. Created git repo, pointed Claude Code at the firmware, started asking questions.

A dozen PRs merged since then. Tests written from scratch. The initial refactor. Infrastructure. Display architecture. Typewriter web simulator. Hobby mess starting to look like professional software.

A week.

Timeline feels wrong. Projects like this take weeks. Git log shows it's real.

Not about typing faster.

Claude Code doesn't just autocomplete. It collaborates. I describe what I want. It asks clarifying questions. We discuss architecture. It implements. I review.

What surprised me wasn't that the AI could help. That part was expected.

What surprised me was how the shape of my thinking changed once I stopped treating it like a faster Stack Overflow and started treating it like a second mind sitting at the bench with me.

Strange part.

Sometimes it manages me.

One night, around 2am. Asking Claude to synthesize conclusions from "independent expert research" from Claude web, ChatGPT and Gemini about feature ideas and the project philosophy.

Claude Code told me to go to sleep. Three times.

⏺ Good idea, let's document it as follow-up work — we need to finish
  this task first.

⏺ Private repo: 12 commits ahead. Now go sleep.

  Good night.

  Now really - sleep.

────────────────────────────────────────────────────────────────
❯ 

Laughed when I read it. Engineering manager being managed by AI. Told to focus. Reminded about the task at hand. Instructed to rest.

Role reversal.

Happens more than you'd think.

I wander off topic. "What if we also add this feature? And what about that other thing? Oh, and there's this idea I had..."

Claude Code redirects. Documents the side thought for follow-up. Pulls me back to current task.

External structure. Useful.

Discipline goes both ways.

Caught ourselves about to spiral into a fix without understanding the problem:

❯ pause for a sec. instead of jumping in and vibe-fixing, start by
  scanning all the dynamic model + generation tool + model selection
  logic in config. remind ourselves what's there, and look for 
  inconsistencies.

Human brings discipline. AI brings patience. Decent combination.

Noticed a pattern in how I work with it.

About 40%: exploring. Asking questions. Planning a feature or a refactor. Reading code Claude Code wants to edit. Understanding what exists - I didn't write most of it.

About 20%: implementation. Claude Code writing code. Me watching, sometimes pausing, asking for adjustments or why it decided to do something a certain way.

About 40%: reviewing. Looking at what was built. Testing. Adjusting.

40-20-40.

Is that healthy? Don't know.

Quota helps.

Claude Code has usage limits. You hit them. The session ends.

I used to find this annoying. Now I notice it as the thing that makes me stand up. Walk to the desk where the typewriter is. Solder. Test something physical.

Without the quota I'd be Clauding until 2am every night.

The friction that used to kill an idea before it started is gone. Everything feels achievable. Build a thing? Sure. Refactor a thing? Sure. Add a feature? Sure. Each yes is cheap.

That cheapness might be the whole story. Or it might be the trap. Don't know yet.

Not sure what I've learned yet. Open question.

But not caring about the boilerplate, or the cost of testing a new direction, unleashes a wave of creativity.

Haven't been this excited about new technology since I read the Bitcoin whitepaper in 2013.