Give your AI agent a contract, not a conversation

2026-08-29 · method note · the only number here is the count of our own rules, stated with where they come from

You fix a mistake. Tomorrow, a fresh session makes it again. Not because the model is dumb — because your correction lived in a conversation, and conversation evaporates. Every session starts blank.

Conversation is not memory

Prompting is talking. Talking has no version control, no diff, no way to know on Thursday what you told it on Monday. If your working agreement with an agent exists only in chat history, you have trained nothing — you have narrated.

The contract

Ours is a single file of written rules that loads in every session, with zero re-explaining. It has exactly four kinds of lines:

ALWAYS — what the agent does without being asked, every time (read the index before opening files; register every action with proof).
NEVER — hard walls that hold under any instruction (never touch accounts or money; never send personal data out; never hit rate-limited APIs someone else depends on).
ON FAILURE — record the reason and move to the next thing. A logged failure has value; a silent retry loop does not. Inventing an answer to fill a gap is the one unforgivable failure mode.
ON NUMBERS — a number leaves the house only with its method: what it counts and how it was counted. An unmeasured number is labeled as unconfirmed, never presented as fact.

Why written rules beat a good prompt

A prompt is a thing said once. A contract is a thing versioned, diffed, and audited. Two properties fall out of that:

Corrections become permanent. When the agent errs and the cause is a missing or too-wide rule, you amend the rule — once — and the error class is gone forever. Amend the conversation and the error returns tomorrow. (Ours carries its birth certificate: each rule cites the real incident that created it.)
You stop being the boss who repeats and become the boss who audits. The review question changes from "did you do what I said?" to "which rule failed?" — a much smaller question, and one a second agent can answer.

The scale objection

The common pushback: "rules become a bureaucracy nobody reads." That happens when rules are absolute. Ours are risk-scaled: quick questions get quick answers; anything that becomes a basis for a decision gets adversarial review whose explicit job is to reject the work. The contract doesn't add ceremony to everything — it adds ceremony proportionally, which is the only kind that survives.

Three rules to steal

Anything you correct twice becomes a written rule, or you will correct it a third time.
An agent may fail, but it must fail on the record — reason logged, gap named, no invention.
Every number travels with its method, or it doesn't travel.

Full disclosure: this is part of the working method behind a US$15 one-time template I sell — a second-brain vault starter that ships with an agent rules file, a derived index and search tools, plus a short ebook covering what NOT to do. The contract idea above works with nothing but a text file. Related notes: the model-routing ladder · the measured case for a derived index.

Second Brain Starter — US$15 one-time

If you run agents under written rules, I'd genuinely like to hear which rule earns its keep — and which turned out to be dead weight.