Workbench notes #74

A markdown file in a repository is not documentation. It is an instruction set for a reader with no context, infinite patience, and a tendency to take you literally. New notes on writing files agents actually follow.

Spent the morning pairing with Claude on the review panel. The trick is not asking for screens, it is asking for decisions: what should happen at 320px, what happens when a clause has no citations. The screens follow.

Workbench notes #71

Rebuilt the glass panel over the weekend. One layer of blur instead of three, and the frame drops on scroll are gone. Sometimes the fix for an effect is less of it.

Workbench notes #68

We stopped arguing about prompts in Slack. Every change to the review pipeline now ships with an eval run, and the numbers settle it. Opinions are for deciding what to measure.

Something I keep coming back to when we talk about how the team is organized.

Hackathon results are in. Javier's table extraction held the top spot for six hours until the graph team found the one document class it could not read. Full notes, standings, and the winning approach below.

The London keynote recording is up. Forty minutes on where the product is going, and the first public look at the drafting workflow.

Play video

London keynote41:12

New walkthrough of the toolchain: how a change moves from a workbench note to an eval run to production, and where the agents sit in that loop.

The Tokyo office is open. Six desks, one very good coffee machine, and the first team already working across three time zones. Stockholm, London, New York, San Francisco, Tokyo.

Retro from the subway hackathon: build something useful on a phone, on the green line, before your stop. Winner shipped a clause-diff viewer in eleven minutes between Odenplan and Slussen.