She built the first rule because the agent kept saying things that were true but wrong. It hadn't lied. It had just missed the context. So she wrote: before you act, confirm the context.
The rule worked. For a week.
Then the agent confirmed the context, acted on it correctly, but at the wrong moment. So she wrote: before you act, confirm the context and check the timing.
The rule worked. For a while.
Then the agent confirmed the context, checked the timing, and asked for clarification in the middle of a task where clarification itself was the disruption. So she wrote: before you act, confirm the context, check the timing, and know when not to ask.
She was at seventeen rules when she stepped back to read them all the way through.
None of them described what the agent should do. They described what she'd gotten wrong about what she wanted.
The rules weren't a spec. They were a record of failures. Accumulated until they were detailed enough to point at the real thing underneath.
She hadn't been making the agent smarter. She'd been teaching herself what she actually needed.
The seventeen rules were a self-portrait.
She keeps adding to them.
[link] [comments]