Site Navigation

Putting Claude on Rails

Introduced by: Pete Kaminski
Depth: Deep thread

Pete's core methodology for working with Claude Code, and the central teaching of this conversation. The metaphor is literal: you build railroad tracks (a project plan) before the train (Claude) starts moving. Without rails, the train barrels down random paths, crashes, and burns tokens.

The Conversation

Pete introduced this concept by contrasting his approach with Dave's:

Pete: "So, the way I've been telling people this, my brother Dave is going great guns with Claude Code, but I didn't get to teach him how to do it, and he's just doing it out of his head, right? So he ends up with Claude Code, barreling down, like, it's a freight train, barreling down something, right?"

Pete teaches his PKAI Course students to always start with structure:

1. Make a new project folder
2. Write your goals in a short statement
3. Have Claude write a project plan to meet those goals
4. Review the plan — look for "kinks in the rail"
5. Fix the kinks before the train starts moving
6. Then say "okay, Claude, do this project"

Pete: "I never start people off with, hey, Claude, do something. It's always write a project plan for the thing that you're about to do, right? And then we'll check it together."

Dave recognized this was the thing Pete had told him about at the very beginning:

Dave: "Oh my god. You probably shared that with me, but maybe not."

Pete: "I forgot."

Dave: "Already that I need a project plan? Because I got so far into it."

The Live Demo

Pete created a Palmer vault in real time, demonstrating the full workflow:

1. Made a folder called "Projects"
2. Wrote a brief goals statement (about 40 words) for identifying Palmer handwriting dates
3. Had Claude expand it into a full project plan — rationale, sub-projects, classification systems
4. Showed how to break a big project into sub-projects: "catalog personal letters," "build a Palmer taxonomy"

The project plan Claude generated was impressive — it even knew about Spencerian vs Palmer distinctions, Zaner-Bloser, and Danealian styles.

Key Points

• Structure prevents the train wrecks Dave experienced
• A project plan costs almost nothing in tokens compared to hours of unstructured wrestling
• The plan doesn't have to be long — 40 words of goals can produce a solid plan
• Review the plan before executing — this is where you catch problems cheaply
• Break big projects into sub-projects; each sub-project gets its own plan
• After execution, write a work log and make a git commit

The Quotable Moment

Dave synthesized the whole philosophy into a single insight:

Dave: "You don't have to have structured data — structure the way you work with it."

Pete realized he'd never said this explicitly to his class, even though it was implicit in everything he taught. See Structured Data vs Structured Interactions.

Related

• Dave's Claude Struggles — what happens without rails
• Structured Data vs Structured Interactions — the insight this methodology implies
• Agentic AI Development Cycle — Pete's visual diagram of the full workflow
• PKAI Course — where Pete teaches this approach
• Three Times Three Thinking — Dave's alternative (less effective) approach to the same problem

---