// talks / spec first: how to collaborate with ai at scale

Spec First: How to Collaborate with AI at Scale.

Writing intent down precisely enough for an agent to execute, and plainly enough for a human to trust — proved four different ways.

July 19, 2026 · ~45 min, incl. 3 live demos
// slides
open full screen ▸

Abstract

I have outpaced my own ability to organize intent, and the gap compounds every week. Thirty-plus repos, tools, and prototypes in every stage of disarray. The old way of keeping it straight — holding it in my head, syncing it in meetings — stops working at that speed.

The fix is to write the intent down as a shared object that runs two loops at once: a local loop with my own agent, and a global loop where other people and their agents sync to what I wrote, with no meeting required.

The talk proves that four ways — scaling up through a chain of specs from SODL to Faraday, out across parallel worktrees, across seven-plus contributors on one spec, and through time as a single spec matured from contract to UX to prototype to evals.

Live demos

Three demo breaks, roughly three minutes each:

  1. The SODL spec deciding what the C parser accepts and rejects
  2. A chief-of-staff skill reporting on real activity across 30+ repos
  3. Handing one spec to a fresh agent session, cold, with no re-briefing
// resources
◂ all talks