Curiosity, made tangible.

Before an agent starts, I ask it to name the result, what would prove it works, and what must stay unchanged. A completion report is a claim to inspect. If a page is supposed to move, I want it opened and used. If a test passes, I want to know what that test actually establishes—and what it leaves unanswered.

My instructions explicitly ask agents to check my claims by the same standard as their own. When a premise is wrong, I want the evidence, its consequence, and a better next step. My confidence, repeated agreement, or the effort already spent should never make an unsupported idea sound more certain. My approval lets work proceed; the underlying claim still needs evidence.

When independent review is useful, I want a fresh reviewer to receive the work and the criteria, with room to find the original approach wrong. Two AIs agreeing is still a judgment to examine. When they disagree, I trace the difference to evidence. A reviewer’s finding has to hold up before it becomes another task.

My setup gives agents a searchable local memory of decisions and corrections, alongside instructions for each project. I ask them to check remembered context against the current source before acting. Handoffs carry decisions, open questions, and supporting evidence. Reusable guidance stays separate from the immediate task, and each read has a purpose.

After a useful correction, I ask for the situation that triggered it, the procedure that worked, and the limits of the lesson. That goes into guidance the next agent can find and use. I also ask for repeated rules to be combined and stale instructions retired. Remembering should make the next attempt better.

Tokens, time, and attention all count. I ask agents to start with the smallest useful outcome, justify additional agents and checks, and name a stopping point. A reversible experiment can move quickly; work affecting someone’s data or a live product needs more scrutiny. I want to judge the effort against a verified result. If nothing holds up, I would rather hear that than get an inflated answer.