AI updates and workflow thinking.

Practical notes on building with AI: what works in production, what breaks, and how to think about it.

Why most AI automations fail in production

The demo worked. The production deploy didn't. Here's the pattern I see repeatedly, and the single process step that prevents it.

A prompt written as a one-off request breaks under variation. A prompt written as an operating rule handles edge cases by design.

Any AI agent touching client communication, money movement, or irreversible actions needs a human checkpoint. Here's how I wire it.

Most automation breaks because the workflow assumes the input. Add a classification step at the front and the whole system becomes robust.

The tools I reach for on real projects: n8n, Claude, Airtable, and a few others. With honest notes on where each one actually fits.

AI is not the answer to every inefficiency. Some processes should stay manual. Here's how I decide which ones.

Get new posts when they land.

No newsletter cadence. Just when there's something worth reading.