I've been building software with AI coding agents for a while now, and the experience has shifted in ways I didn't fully expect. These are my observations—not conclusions, but patterns worth naming. I'll keep adding to this as I go.
Essays
Building with AI
What happens when the feedback loop between idea and working code collapses to minutes? Something fundamental shifts—not just speed, but the nature of the work itself. These are four patterns I've noticed so far.
The four themes:
- The Solo Creator Dynamic — why AI makes development feel like solo creative work
- The Training Density Effect — why AI reliability varies by domain, and how to calibrate trust
- Knowing When to Constrain the Problem — prompt specificity should match your confidence, not the task's complexity
- Discovery Before Prioritization — when building is cheap, the logic of traditional planning inverts
Read the essay →
Deep Dive: The Multi-Agent Loop
Some early thinking on what happens when you stop working with one AI agent at a time -- and what that might mean for how solo builders work.
What's covered:
- The thinking/building split — why separating strategy from execution adds leverage
- How the Qupi product advisor + Claude Code setup works in practice
- Scaling to parallel implementation streams and what makes them work
- What changes about your own role when multiple agents are running
Read the essay →
New
Deep Dive: AI as a Feature
What happens when you stop treating your app's AI capabilities as a black box and start building the tools to measure, compare, and refine them.
What's covered:
- Extracting AI capabilities into standalone, testable modules
- Building an evaluation harness with single-input, batch, and comparison modes
- Context fixtures and match rules for testing non-deterministic behavior
- How "building with AI" and "building AI into the product" reinforce each other
Read the essay →
Deep Dive: Why, What, and How
AI coding agents are remarkably good at figuring out how to build things. But figuring out what to build -- and why -- is a different problem entirely.
What's covered:
- Why agents excel at the how but can't supply the why or what
- The subtle risk of confident-sounding answers without real context
- What you actually need to bring to each development session
- The product advisor setup -- and why a thin version isn't enough
Read the essay →