Notes from building with agents.

• July 2026A Wendy for a team of one

  Hedge funds have performance coaches and AI labs are hiring philosophers — because past a bar, the constraint is psychology, not knowledge. How James Lim's method became a coaching skill inside my projects, the session that got Drafty in front of a real person, and the two fake-progress urges it caught live.

• June 2026I let Claude run a peptide wiki I knew nothing about

  A peptide wiki I never wrote, run to 111K impressions by an autonomous Claude Code loop — the daily loop, the three skills, and the BigQuery setup that makes it work.

• June 2026I mined 224 transcripts to salvage a model too good to be legal

  Fable 5 got pulled three days in. The leverage I'd already banked — a re-architected workflow and an operating manual mined from 224 transcripts — and whether it holds on Opus 4.8.

• June 2026My skills go stale in two months

  A wiki that runs itself, re-skilled with my current copy and design playbooks — and why a skill written down compounds across every project.

• June 2026A cinematic hero for 80 cents

  The real cost of a video model isn't the render or the safety filter — it's iterating, and every pass is billed. How a cinematic Veo 3.1 hero stays at 80 cents: settle composition on free GPT Image 2 stills, iterate motion on Lite, pay full price only for the keeper.

• June 2026Realtime voice on Gemini Live 3.1: 13× cheaper than OpenAI

  OpenAI ships WebRTC and it just works. Gemini doesn't — so here's the realtime-audio setup, end to end, and the two weeks of echo bugs it took to get there.

• May 2026Pick the model for the job

  Claude for code, Gemini for images, Qwen for speech. Routing to six models across five projects.

• May 2026Claude as ADHD assistant

  Concrete examples of using Claude to get a thought out the door when executive function isn't there.