CH 06 · Prompt Architecture · 04/10
Read as one pageWrite rules as consequences, not just actions
For domain agents, the difference between a toy and a trustworthy tool is often in the rules that encode consequences. "Edit the text" is an action. "Any edit that adds or removes a numbered item shifts every following number, so renumber the siblings and update every cross-reference to them in the same edit" is a consequence. The second is hard-won domain expertise expressed as instruction, and it's what makes the agent safe to use on real work. Capture your experts' knowledge as these consequence-rules.
Finished this idea?
Your learning progress stays in this browser.