Revisiting AI-Assisted Development
One Year Later
From "Exercise Left to the Reader"
to Production-Ready Infrastructure
- Intro self & the 'Robot Terror' side project.
- Context: Replaced manual work with AI in 2026.
- Source: Recent blog post review.
The Starting Point (2025)
├── handler.py
├── api-gateway-how-to.md
└── README.md
Great Code.
Manual Infrastructure.
Setup was a 60-step process.
- 2025: "TerrorTunnels" API.
- AI wrote the Python code perfectly.
- BUT: Creating the API Gateway, IAM, etc. was manual.
- "Powerful code generation, wrapped in manual toil."
The "Last Mile" Problem
Translating code implementation into cloud reality was the bottleneck.
- The gap between `handler.py` and a working URL.
- AI could *tell* me what to do, but couldn't *do* it.
- Result: Friction. Ideas died in the "deployment" phase.
Enter 2026: Agency
CLAUDE CODE
The shift from "Chatbot" to "Agentic CLI".
- Returned to the project in Jan 2026.
- Used Claude Code (CLI agent).
- Different interaction model entirely.
One Conversation
> AWS API Gateway implementation.
Explored codebase. Chose SAM. Generated Template. Fixed Errors.
- I gave a high-level goal, not a snippet request.
- It acted autonomously: Exploration -> Decision -> Execution.
- It even diagnosed its own deployment errors (IAM logs).
The Result
├── template.yaml (Infra)
├── Makefile (Ops)
├── events/ (Tests)
├── samconfig.toml
└── handler.py
Infrastructure as Code
• Versioned
• Reproducible
• 1-Command Deploy
- From manual markdown to full IaC (SAM).
- Added tests (Makefile, curl harness).
- Total transformation of project maturity.
Abstraction Shift
2025
Knowledgeable Colleague
"I can't touch the keyboard."2026
Junior Engineer
"I'll handle that task."- 2025: I was the executor. AI was the advisor.
- 2026: AI is the executor (within guardrails).
- The shift is Agency.
The Expertise Paradox
As tools become more capable,
Domain Knowledge matters MORE.
• Design Judgment: Is this architecture right?
• Security: Are IAM roles scoped correctly?
• Ops: What will this cost?
- Delegating execution requires ability to Review.
- I couldn't verify the SAM template if I didn't know CloudFormation.
- Trusting requires validation.
The Trust Gradient
Volume of delegation increases. Review remains constant.
- Gradient has shifted up.
- Result: Lower "Activation Energy" for side projects.
- Ideas become working systems much faster.
AI Amplifies Expertise
It doesn't replace it.
The Sweet Spot:
Powerful enough to delegate.
Constrained enough to verify.
- Infrastructure as Code that I can trust.
- Not bad for an afternoon's conversation.
- Q&A.