os

My product judgment, written so AI agents can run it, and checked monthly to see that they still fire.

I run partner success in adtech, getting enterprise partners from "interested" to live, and I build most things through AI agents. Early on, the agents' failures were judgment failures: a polished partner email that asked for nothing, a feature green-lit because one loud partner wanted it.

So I wrote my judgment down as lenses: short rule files an agent reads before it acts. This repo is the public part of that system.

See it work. Four of the five lenses were tested the same way: two fresh agent sessions, same task, one with the lens and one without.

• Taste: a partner change notice
• Commercial: a renewal that asks for a worse split, a guarantee and exclusivity
• Product: the biggest partner demands a custom feature in six weeks
• Motivation: an upgrade email that needs 30 of 40 partners to act

Lenses

Lens	Runs when
taste	Last, before anything outside the team sees an artifact. Returns ship, ship with fixes, or not yet, with the why and rewritten text.
product	Before committing people or money, saying yes to a partner ask that implies build, or choosing between options. Returns build, configure what exists, test first, or decline.
commercial	Before agreeing to a rate, split, guarantee or exclusivity. Returns a position with a floor, a walk-away, and an ask for every concession.
motivation	Before anything that needs a specific person to act: a change notice, an adoption plan, a buyer pitch. Predicts whether they will, and what to change.
anti-ai	Before shipping AI-assisted code, decks, docs or analysis. Catches the tells writing checkers miss.

Each lens is a markdown file with skill-style frontmatter. Drop it into a skills folder as taste/SKILL.md, or paste it into any agent's instructions.

How the full system works

1. Lenses. Short files of numbered rules, each with a stated trigger: product, motivation, strategy, taste, commercial, design and more. Five are published here.
2. Registry. One file lists every lens and the situation that fires it, and nothing else lists them. I learned that rule the hard way: twice, a hand-written list went stale and lenses silently stopped firing.
3. Hooks. Scripts that turn the most important rules into mechanisms: a content-safety scan before any push or PR, a design gate before the first UI edit, a block on pushing to main. One is published as an example: enforcement/no-ai-artifact-md.sh.

Testing

Prompt systems rot quietly. Someone rewords a trigger, a real task stops reaching the lens it needs, and nothing errors.

Routing regression. Nine real tasks run through fresh headless agent sessions. Each run checks which lenses the agent opened, and which it shouldn't have. product, motivation and taste are published; the rest are in my private set.

Task	Must open	Must not open
Tighten a partner email before sending	motivation, taste
Partner asks for a custom ranking feature	product, motivation
Should we build a self-serve preview tool?	product
Positioning doc for a new partner segment	strategy, retail media
Coach a direct report on discovery-call notes	motivation, taste
Pressure-test an SDK placement idea	product, taste
Update a README		taste
Build a new UI component	design, workflow gates
Open the PR for a UI change	design de-slop check

Last run, 2026-10-11: 8 of 9 pass. The miss: asked to open the PR for a UI change, the agent ran the code and safety checks but never reached for the AI-slop design check on its own. Written rules alone missed it. In my live setup, a ship gate blocks that push until the check has run, and that's why the gate exists.

Monthly audit. Automated checks for broken cross-references, unregistered lens files, hooks pointing at missing scripts, the same principle restated in two places (which is how drift starts), and whether the workflow gates are actually followed.

Most of those checks exist because something broke without anyone noticing. In one case, a hardcoded file list let two new lenses go unchecked for weeks. The fix was to derive the list from the registry so it can't be hardcoded again.

Deliberately not here

• Employer and partner context. Partner names, revenue numbers and internal schemas stay private. The lenses work without them.
• Most of the enforcement layer. The hooks are wired to my machine and accounts. One sanitized example is included.