Notes from the workshop

What we learned building things, written up while it was still fresh. No thought leadership, no listicles.

Writing acceptance criteria for things that are probabilistic

A feature that calls a model does not have a single correct output to test against. That does not mean it cannot have acceptance criteria — they just look different.

TestingAIQuality

Quality · 6 min

14 Jul 2026

The audit we run on our own projects every quarter

We turned the checklist we use when reviewing other teams' systems inward, on our own delivered work, and found exactly the kind of drift we warn clients about.

ProcessQualityEngineering

Process · 6 min

13 Jul 2026

What an agent gets wrong about database migrations

Coding agents write plausible-looking migrations quickly. The mistakes cluster in a small, predictable set of places, which makes them easy to check for specifically.

AIEngineeringData

AI · 6 min

10 Jul 2026

A definition of done that includes deletion

Feature flags, migration shims and temporary compatibility code accumulate quietly. We started treating their removal as part of the ticket, not a separate one nobody files.

ProcessEngineeringDelivery

Process · 5 min

09 Jul 2026

The difference between a model agreeing and a model being right

Ask a model whether your approach is sound and it will usually find a way to say yes. That tells you almost nothing, and we had to build a habit around it.

AIClaudeEngineering

AI · 6 min

08 Jul 2026

Build note: a translation layer for a supplier feed nobody could change

A retail client's stock feed arrived in a format from 2009, undocumented, from a supplier with no incentive to update it. The interesting engineering was all on our side of the wall.

Case studyEngineeringData

Case notes · 6 min