/writing

Why scope creep kills more zero-to-one products than any rival, and what to cut first.

The two voices a founder writes in, why the first matters more early on, and how to tell which one is showing up.

Three small loops to test demand before writing a line of code. Run them in a weekend.

Why the rental row never had a status column, and what we built instead.

Combined transaction tables look right on day one and are unrecoverable on day three hundred. The schema choice that makes audits a query.

How the console became a system of display, not a system of record - and why that decision held up under load.

Capture the consultation as data, not as the call. The discipline behind a telemedicine app that worked when the network didn't.

Why we abandoned OR-Tools for a deterministic algorithm operators could read. Trust beats optimality for local services.

Buy the transcoding stack. Build the reliability layer above it. Make audio-only first-class. The notification is the product.

Brackets aren't a rendering problem. They're a state-and-event problem with a particularly visual presentation.

Why the unit of billing was 'active rental day,' not 'monthly subscription' - and how that single choice rewrote the architecture.

Recurring donors don't churn loudly. They quietly stop trusting the platform. Here's how the product fought that decay.

Most B2B SaaS for small businesses fails the same way: it treats the owner like an analyst when they need an executor.

Speed without breaking trust. The operating constraint when a healthcare app has to ship in days, not quarters.

Most marketplaces eventually become operations platforms. This one started as one - and that was the right call.

Designing for an audience the tech industry doesn't usually design for - and what changes when you do.

Local tournaments don't need ESPN. They need the WhatsApp group of 80 organizers and players to stop arguing about the bracket.

The first 20 turns felt like magic. The next 10 felt like a different model. Here's what was actually happening to my context window.

The agent finishes the task, then keeps going - refactoring three files I didn't ask about. The fix wasn't in the prompt. It was in the harness.

I blamed the model. Then I read my own tool's JSON schema. It was the bug.

Two of the sharpest teams in the field disagree out loud. After trying both, here's what I'd actually code on Monday morning.

Four ways to codify the same recurring thing. I picked wrong, twice. Here's the decision tree I wish I'd had.

Every hook I write is a vote of no-confidence in the model. The question is which votes pay off.

I deleted one memory file and the next session got better, not worse. That bothered me enough to write this.

Same model, different feeling. Same task, different brain. The harness isn't infrastructure. It's the product.

A while-loop, an LLM, and a tool registry. What could go wrong? Most of it.