Teaching Claude to Code: Building a Virtual Software Team from Scratch
Building Sunday Planning: A Founder’s Journey Into AI-Assisted Product Development (Part 2)
About 3 months into the process, Sunday Planning shifted from “maybe one day I could build this” to “I’m actually building this—with AI functioning as a real engineering team.”
And the breakthrough didn’t come from “AI writing code for me.”
It came from realizing:
You can’t ask AI to build an app that has never existed. You have to teach it what you want first.
The First Big Lesson: AI Needs Context, Not Commands
Claude Code didn’t simply ask, “What do you want the app to do?” It asked for:
the vision
the constraints
the feature list
the user goals
technical preferences
design principles
business context
Once I fed it all of this, the conversation transformed.
Claude produced:
a refined tech stack
a mapped-out development plan
initial feature specs
early user flows
architectural choices
even potential pitfalls
It was like talking to a CTO who never got tired, never got defensive, and always answered with explicit reasoning.
Data First: The 40-Table Turning Point
Before we wrote a line of production code, Claude pushed me to design the data model.
I started with a simple schema sketch—12 tables, nothing complicated.
But then I asked:
“Where are the gaps?”
Claude answered with the kind of rigorous scrutiny you’d expect from a senior data architect:
What if a trip has overlapping days?
Do events need time zones?
Should guides have versions? Comments?
How are shared lists tracked?
Should places sync with external APIs?
The schema ballooned from 12 tables… to over 40.
And yet none of it felt excessive. Every table represented a real edge case or future capability that “future me” would thank “present me” for anticipating.
This is when I internalized a key truth:
AI doesn’t remove thinking. It amplifies it.
A Multi-Model Peer Review
Once the schema and plan were drafted, I ran a three-model experiment:
Upload everything to ChatGPT
Upload everything to Co-Pilot
Upload everything to Gemini
Then I asked each one to critique the work:
What’s missing?
What’s flawed?
What’s unnecessary?
What will break later?
I took all their notes back to Claude and said:
“Review these critiques, score their validity, and propose adjustments.”
This became an incredibly powerful workflow.
Not because AI is infallible—but because different models saw different blind spots.
The result was a stronger, cleaner, more robust foundation for Sunday Planning.
The Hidden Enemy: The Context Window
Everything went smoothly—until it didn’t.
Claude slowed down. It would forget details. It refused to continue responses. Even simple prompts would immediately hit token limits.
I eventually discovered why:
Every message sends the entire chat history.
I had overloaded the conversation.
The fix was architectural:
separate conversations
modular files
clean handoff docs
.claude folder for all agent rules
smaller, scoped prompts
explicit memory boundaries
Once I reorganized the system, velocity returned instantly.
Building a Virtual Engineering Team
At this point, I wasn’t working with “an AI.”
I was working with a team.
We created:
a platform engineer
multiple feature engineers
a schema specialist
an integration engineer
a tester
a lead engineer (Claude itself)
Giving them names like Peter (Platform) and Tina (Trips) wasn’t just whimsical—it made communication clearer. Each agent had:
ownership
scope
rules
context files
constraints
design principles
This dramatically improved consistency and reliability.
Finally Writing Real Code
When we finally began building features, something surprising happened:
I wrote far less of the code than I expected.
My role evolved into:
defining the feature
clarifying edge cases
resolving conflict
reviewing drafts
running kickoffs
validating branch outputs
providing product context
ensuring consistency
This is where my product management background became absolutely essential.
Because even with an army of AI engineers—
someone still needs to define the strategy, the requirements, the priorities, the customer value, and the quality bar.
AI dramatically increases speed,
but product management remains the steering wheel.
The New Reality
By this point we were shipping early features:
saving places
planning trip days
writing travel guides
syncing data
testing multi-branch changes
I was no longer learning how to build software.
I was learning how to lead an AI team.
Read more about how AI sparked the journey in Part 1.
In Part 3 (coming soon), I talk through how this changes everything about the future of building products.
