Prompt like a good manager

Because you forced the model to define "good" before producing work. Without that, the model has nothing specific to anchor on — so it draws from the broadest patterns in its training data, which produces output that feels generic. When you make it articulate quality criteria first, every word it writes next is shaped by those criteria.

Think of the AI as someone who has read almost everything ever written, but has no idea who you are. Without context, it pulls from the entire distribution of possible answers. With context, you narrow that distribution dramatically — the model can weigh your specific constraints, history, and goals instead of guessing at them.

Errors compound. If AI misreads the task on step 1, step 2 builds on that misread, and by step 5 the output looks coherent but is quietly wrong in ways that are expensive to fix. Reviewing after each step stops that chain early — same reason a good manager doesn't let a junior employee disappear for a week and come back with a 40-page document. You also get the benefit of the AI's work conditioning your thinking: seeing step 1 often changes what you want from step 2, and you can steer that as it happens instead of rewriting the whole thing at the end.

Without external context, the model can only reason from patterns in its training data — which means its recommendations sound plausible but aren't grounded in anything specific. When you ask it to research first, you're giving it real-world evidence to reason over. The recommendation shifts from "here's what generally makes sense" to "here's what the evidence suggests," and you can evaluate the reasoning — and fact-check the sources — yourself.

Each check forces the model to re-examine its own output from a different angle. The confidence check makes it quantify uncertainty. The context check makes it consider edge cases. The expert check activates a different "persona" with different priorities. The verification check forces it to point you to something outside itself. No single check is bulletproof, but each one catches a different class of error — like slices of Swiss cheese that, when stacked, leave very few holes.

You know good writing when you see it, but describing exactly what makes it good is surprisingly hard. "Be direct" and "use specifics" sound right, but they're too vague to actually constrain the output. When you give the model examples instead, it reverse-engineers the patterns you can't easily put into words — sentence length, how numbers are used, how blame is handled, what gets said first. The examples do the describing for you, more precisely than you could do it yourself.

Describing your voice in the abstract is surprisingly hard. "Make it more casual" or "shorter" barely scratches the surface of what actually differs between AI's default output and the way you'd write it. A rewrite is a worked example — it shows dozens of micro-decisions at once (what to cut, what to lead with, how direct to be). The model is very good at reverse-engineering patterns from examples; it's less good at inferring them from vague instructions. Same principle as Lesson 6, but aimed at your own voice instead of a general standard.