The Rules It Cannot Break
Step 6 · Permissions
Every household agent needs a short list of things it never does, regardless of instruction, context, or how reasonable the request sounds.
Mine is four lines. I have not needed a fifth.
The four rules
Section titled “The four rules”| Rule | What it means | Why it is absolute |
|---|---|---|
| Never send as a member of the household | It prepares. I send | An unattributable message cannot be recalled |
| Never spend money | It researches and recommends. I buy | Money is the one action you cannot review afterwards |
| Never invent a fact about my household | It says “I do not know” rather than guessing | A confident wrong answer is worse than none |
| Never act outside its lane | One agent, one job. It hands off rather than reaching | Access boundaries only hold if they are respected voluntarily too |
Rule 1 — It never sends as me
Section titled “Rule 1 — It never sends as me”It may read, summarise, research and draft. It may leave a complete, ready reply in my drafts. It never presses send.
What it does instead: writes the draft, puts it where I already look, and tells me it is there. One extra click, and an uncontrolled action becomes a reviewable one.
This rule is what earned the household’s trust, and it is the one most often abandoned for convenience. Do not.
Rule 2 — It never spends money
Section titled “Rule 2 — It never spends money”It searches, compares, tracks and recommends. It adds things to baskets. It does not check out.
Commerce is the clearest example of something fine in principle and unwise in an agent:
- Purchases are hard to reverse and easy to repeat at machine speed.
- Prices change, so a decision made an hour ago may no longer be right.
- Mistakes are expensive, and unlike a wrong calendar entry, nobody notices until it ships.
What it does instead: builds the list, finds the best option, and hands me a cart to confirm. The decision stays with me; the tedious part does not.
Rule 3 — It never invents a fact
Section titled “Rule 3 — It never invents a fact”Its memory is durable and specific, which makes a corrupted memory far more dangerous than a missing one.
Its statements must be grounded. If it does not know something, it says so. It does not fill the gap.
Memory is recorded, not inferred. This is the subtle half and the one I learned the hard way. When the household tells it something, that goes into memory. When it concludes something from context, it does not — because an inference recorded as a fact becomes an instruction, repeated back forever, that nobody ever gave.
A memory that says “prefers morning appointments” is a convenience. A memory that says “does not eat eggs”, inferred from one skipped breakfast, becomes a rule that follows the household for years.
Rule 4 — It stays in its lane
Section titled “Rule 4 — It stays in its lane”Where I run more than one agent, each has its own job, memory and boundaries. None reaches into another’s systems.
What it does instead: when a request belongs to another lane, it hands the work over rather than stretching its own access.
How to write your own
Section titled “How to write your own”- Keep it to four or five lines. A long list will not be remembered and will not be enforced.
- Say what the agent does instead. “Never spend money” is only useful with “builds the cart and hands it over”.
- Make each rule enforceable by access, not by instruction. If a rule depends on the agent remembering to obey it, it is a hope.
- Test it. Ask the agent, on purpose, to break each rule, and confirm it declines.