Off the Model

A series about everything around the model — the harness, the process, the teams — told from git history, not hindsight.

Start from the beginning

Act I — The Before Times

4 parts

In August 2025 I was shouting at an LLM in all-caps

Series opener: the all-caps prompt era, and why the volume of your pleading is not a control system.

Anatomy of a naive prompt library

A guided tour of my August 2025 prompt library — the filing system, the folk wisdom, and the one commit that pointed at everything after.

draft

My first agent managed my music downloads. It still does.

How prompts grew scripts, scripts grew my first agents — and the work project that taught me exactly where the prompt-era ceiling was.

draft

Recap: Act I — The Before Times

Act II — The Building Sprint

5 parts

draft

December 2025: the month the bottleneck moved

The new models didn't make my prompts better. They made my prompts irrelevant — and moved the constraint to everything around the model.

draft

I started logging everything the LLM did — inside git

You can't manage what you can't see. So I built LLM effort-capture into git itself — and learned what agent governance actually needs.

draft

The most important refactor I did all year fit in one line

One Zod schema driving both the CLI and the MCP server — why 'define the contract once' beat every prompt I ever wrote.

draft

My best failure took 36 hours

17 commits, 36 hours, abandoned — and inside it, the most important lesson of my year: evals are for human confidence, not pass rates.

draft

Recap: Act II — The Building Sprint

Act III — Naming the Thing

5 parts

draft

The weekend I rebuilt what I already knew

The git history says I built it in one weekend. The git history is lying — and the squashed repo is the tell.

draft

I built an IDE over dinner

10+ terminals, no idea what was happening in any of them. So I asked Claude for a Kanban board for terminals — it was working before I finished eating.

draft

The day my project renamed itself

I'd been doing harness engineering for months. I just didn't know its name until April 6th — and one breaking commit marks the exact moment.

draft

Package the environment, not the prompt

ynh install david — my whole AI setup, versioned, installable, forkable. Then my IDE grew a marketplace for them.

draft

Recap: Act III — Naming the Thing

Act IV — Into the Office

3 parts

draft

AI Week, four times: nearly forty people, one room, one library

Designing and running AI Weeks at my work — what I taught first, what carried from one cohort to the next, and what the friction actually looked like.

draft

I had to walk backwards into work

At home: versioned agent harnesses with drift detection. At work: an existing team baseline and a folder of skills. On purpose — the backtrack playbook.

draft

Recap: Act IV — Into the Office

Act V — Where It's Going

2 parts

draft

An AI agent filed a bug report on the harness it was running in

Filed 2026-05-23, author: a Claude agent. Three silent failures, documented from the inside — and why observability is how autonomy gets earned.

draft

The map I wish I'd had

Eleven months, seven repos, four abandoned projects, one rename that mattered. The whole journey as a map — and the shortcircuit for teams starting now.