Code with Claude London 2026: Notes from Two Talks

Recently I found some time to watch two talks Anthropic gave last week at Code with Claude London 2026—one was Margot’s Prompting Playbook, and the other was Legora sharing their experience with legal AI. Here are a few notes.

An infographic illustrating a three-stage “Generate, Evaluate, Repair” workflow, with a little robot racing down a track, symbolizing how a small model catches up to a large one through good workflow design

Margot’s Talk: Small Models Can Do Big Things Too

In Margot’s talk, what stuck with me most was one data chart: using Sonnet paired with this three-stage “Generate, Evaluate, Repair” workflow, the pass rate actually matched that of Opus with extended thinking—yet it cost only half the tokens and half the time.

In other words, as long as you design the workflow well, a small model can do big things just the same.

Legora’s Talk: Legal Work and Software Engineering Are Similar at Their Core

Legora’s talk centered mainly on contract scenarios (to be honest, the Swedes’ English was a bit hard to follow), which doesn’t have much to do with the patent-response work I do.

But one line Jacob said at the start really resonated with me: “Legal work and software engineering are similar at their core—you can just borrow their best practices directly.” That line lines up quite well with how I currently see things.

Planting a Seed

Maybe some of the ideas I picked up can’t be put to use right away, but it’s like planting the seed of a phoenix tree—perhaps at some moment in the future, it will sprout.

For anyone interested, you can find these two talks on YouTube and give them a watch.