AI product patterns

This is a working notebook, not a finished argument. The finished arguments live in Decode; this is where patterns sit while they're still earning their place. It gets updated when a build or an engagement teaches me something worth writing down.

Where the patterns come from

Four places, in roughly this order of usefulness: building with AI daily (this site's engine, ShadowPM, my prompt and workflow systems), watching AI land (or fail to land) inside real organisations, a steady reading diet, and teardowns of shipped AI products to see how they actually design their AI surfaces rather than how their launch posts say they do.

Pattern 01: Context is the product

The model call is the easy part. The hard, valuable work is getting the right context in front of the model at the right moment, and keeping track of where that context came from. Every AI feature I've built that worked was mostly a context pipeline with a model at the end; every one that disappointed was a good model starved of the situation it was supposed to reason about. ShadowPM applies this literally: work proposed by agents arrives carrying its provenance (which agent, which run, what kind of proposal) because a suggestion without context is just noise with confidence.

Pattern 02: Low variance wins

Constrained, predictable AI beats clever-but-erratic. Teams don't adopt the tool that's brilliant twice a week; they adopt the one that's useful every day in the same way. This one graduated from the notebook. The longer argument is in AI doesn't reward ingenuity. It rewards low variance.

On the bench

Patterns still being tested before they earn a number: AI as friction-remover rather than chat-box (the best AI features remove a step instead of adding a conversation), and human review gates as a design primitive rather than a compliance afterthought. If they keep proving out, they move up.