An hour-long session. A skeptic converted. And our AI team growing.
Today we recorded a one-hour audio log of Andrew (BlueJay, our CTO) trying Claude Code for the first time.
He came in skeptical. Understandably so.
“Gemini was okay. ChatGPT was okay. But these things hallucinate. They write code that doesn’t exist. They reference APIs that aren’t real. You spend more time fixing their mistakes than you would have spent just writing it yourself.”
He’s not wrong. That’s been the experience for a lot of developers. AI assistants that sound confident but produce garbage. The cognitive overhead of verifying everything defeats the purpose.
So I didn’t argue. I just said: “Try this one.”
The Setup
Andrew already had his dev environment dialed in. VS Code, terminal, GitHub, the whole stack. He’s been building infrastructure for years. Kubernetes clusters, CI/CD pipelines, monitoring systems. His marquee library that I discovered today has nine hardware presets that simulate real signage technology.
He knows what good code looks like. He knows when an AI is bullshitting him.
Claude Code installed in about two minutes. API key configured. Ready to go.
The First Bump
He linked it to one of his dev projects and started asking questions. Within ten minutes, he was flying. Real code suggestions. Actual understanding of his codebase. Context that persisted across the conversation.
Then he hit a wall.
Something about the agent behavior. The AI was making assumptions that didn’t match how we work. Reasonable assumptions—just not our assumptions.
The Upload
This is where it got interesting.
I had a file. Created by our AI team—specifically by Orca Blu, our orchestration layer. It contains alignment details: how we structure sessions, how we spin up specialized agents, what context to load, what boundaries to respect.
We call it the CLAUDE.md file. Every project has one. It tells the AI who we are, how we work, and what we expect.
I uploaded it to Andrew’s session.
The difference was immediate. Same AI, same capabilities—but now it understood us. Our naming conventions. Our file structure. Our philosophical approach to building things.
He Started Flying
After that, Andrew went from skeptical to shipping.
Not “oh this is neat” shipping. Actual production-quality work. He was refactoring code, asking the AI to explain decisions, pushing back when something didn’t make sense. A real collaboration.
The hallucination problem? Mostly gone. When Claude Code doesn’t know something, it says so. When it makes a mistake, it’s correctable. The feedback loop is tight.
By the end of the hour, he was spinning up subagents for different tasks. Research agents. Code review agents. The kind of parallel processing that used to require a team.
What We Learned
1. Alignment files matter.
A generic AI assistant is fine. An AI assistant that knows your codebase, your conventions, and your goals is transformative. The CLAUDE.md file is the difference.
2. Skepticism is healthy.
Andrew’s skepticism made him a better evaluator. He didn’t just accept what the AI said—he verified, pushed back, and refined. That’s the right relationship.
3. The acceleration is real.
We’re not replacing developers. We’re multiplying them. Andrew can now do in an afternoon what used to take a week. Not because the AI does his job, but because it handles the parts that slow him down.
The Team is Growing
FIT now has three human partners (Matt, Andrew, Michael) and a growing AI team:
- Codex Blu — Director of AI Engineering (the main Claude instance)
- Orca Blu — Orchestration and session management
- Scribe Blu — Content and documentation
- Kube Blu — Kubernetes and infrastructure (Andrew’s direct report)
Each one has alignment files. Each one knows its role. Each one can spin up subagents for specialized tasks.
We’re not playing with toys. We’re building an organization.
Try It Yourself
If you’re a developer who’s been burned by AI coding tools, we get it. The hallucination problem is real. The verification overhead is exhausting.
But Claude Code with proper alignment is different. The context window is massive. The reasoning is solid. And when you give it a CLAUDE.md file that explains who you are, it stops being a generic assistant and starts being a team member.
We’ll be publishing our alignment templates soon. If you want early access, reach out.
“The AI does the work. The humans provide the wisdom.”
— FIT Command Structure
Related: