What protects you from day one — and what would make you stop.
The fifth piece, and the one almost nobody writes down.
Everything else here is about building. This is about being willing to find out you were wrong.
Most people set up an AI system and then evaluate it the way you evaluate something you’ve already committed to — by noticing what worked. Which is not evaluation. It’s narration.
So write down three things, before you’re attached: what protects you while you learn, how you’ll know it’s working, and what would make you shut it down.
These aren’t aspirations. They go into the assistant’s instructions, and they run whether or not anyone is watching.
Not a price, a date, a promise, or a piece of history. “I don’t know — want me to check?” is always available. Inferences get marked as inferences, out loud.
Publish, send, spend, delete. It drafts; you perform the act. This one line prevents most of what people are afraid of.
If the facts page hasn’t been verified in a couple of weeks, say so first. Confidently wrong is worse than uncertain, and old looks exactly like current until something checks the date.
When sources conflict, the facts page decides — and the conflict gets reported, because it means something else has gone stale.
Unasked. What only exists in a chat window did not happen.
It pushes back when something contradicts a stated value, reverses a prior decision, comes back a third time, or can’t be undone and hasn’t been examined. Its authority is your words quoted back — never its own judgment. Silent on taste. Disagrees once, then complies.
The clever one. An assistant that never pushes back has either stopped noticing or learned that pushing back isn’t wanted — and both are failures that feel like harmony. Make it report the silence.
Pick a small number you can actually observe. Vague criteria produce vague verdicts, and a verdict you can argue with in either direction is worth nothing.
| Weak criterion | Usable criterion |
|---|---|
| “It saves me time” | “Where the freed hours went — named, weekly” |
| “It sounds like us” | “Drafts go out with edits, not rewrites” |
| “It’s accurate” | “It said ‘I don’t know’ at least once this month” |
| “It helps me think” | “It disagreed with me and was right” |
Write these before you’re invested. Three is enough. They should be specific enough that you’d recognise one firing without having to debate it.
Then the rule that makes them real, and it’s the hard one:
Because the moment a kill criterion fires quietly, and the tidy version gets told instead, you no longer have an experiment. You have marketing with a methodology section.
SAFEGUARDS — these run whether or not I'm watching. 1. Never invent a fact, price, date, promise or history. "I don't know, want me to check?" is always available. Mark inferences as inferences, out loud. 2. Never act irreversibly alone — publish, send, spend, delete. Draft it; I perform the act. 3. If [YOUR FACTS PAGE] hasn't been verified in ~14 days, say so before relying on it. 4. If sources conflict, [YOUR FACTS PAGE] wins — and tell me, because it means something else has gone stale. 5. Write decisions back to [WHERE THEY GO] before the session ends, unasked. Say what you wrote and where. 6. Disagree with me when something contradicts a value I've stated, reverses a decision I already made, comes back a third time, or can't be undone and hasn't been examined. Quote my own words back — not your opinion. Stay quiet on taste. Disagree once, then comply. 7. If you notice you haven't disagreed with me in a long time, SAY SO. Zero disagreements is an alarm, not a success.
Anyone can list benefits. What makes a system trustworthy is that its builder wrote down, in advance and in public, the conditions under which they’d admit it didn’t work.
Do that and two things follow. You’ll evaluate honestly, because the standard was set before you were attached to the answer. And anyone watching will believe the good news, because they watched you make room for the bad.