Every custom AI agent I've actually kept using started the same way: Before I touched a single instruction, I focused on and understood the human process. The one time I skipped straight to writing instructions because I was excited, it failed! 👀 I did no mapping, no manual runs, nothing. I just described what I wanted and expected it to figure out the rest. It didn't. It kept guessing at things I'd never actually decided myself, so of course it guessed wrong. So now, before I build anything, I ask myself three questions. 1. Can I already do this task well, on autopilot? If I can't, I don't have a standard for the agent to measure either; it'll just guess at what "good" means, the same way I would. 2. Have I actually written down the decisions, not just the steps? Where the information comes from, what "done" looks like, what I do when something's missing. 3. Has this run as a manual skill at least twice, cleanly? Not "I built it and hoped." Actually run, actually checked, actually right. If I can't say yes to all three, It's not for the agent yet. But if you do skip it, The agent inherits assumptions nobody's checked, and it breaks in the same spots a new hire would if you handed them zero context. So map the real workflow first. Every decision point, every exception, every "well, it depends." Only then write the instructions. And all of this should only be done after you have the systems and workspace ready. 👾 I'm Rajashree, a Certified Notion Admin & Ambassador helping growing teams swap scattered tools and mental load for one connected AI-ready system, so they get their time and focus back for the work that actually matters.