A four-part recipe for getting useful work out of AI, and the ingredient most people skip.
We’ve been working out how to get genuinely useful output from AI on real production work, not demo-grade output. It comes down to four ingredients, and the one most people skip is the one that makes the rest work.
The recipe
First, type specification. Be explicit about the kind of thing you want out: its shape, its constraints, its format, not just its topic. “Write about onboarding” is a topic. “A 200-word in-app message, second person, one CTA” is a type.
Second, engineered context. Hand the model the material it needs, deliberately assembled, rather than whatever happens to be lying around. Context is something you build for the task, not something you hope is already there.
Third, the human checkpoint. A defined point where a person reviews and corrects before the output moves forward. This is the ingredient that gets dropped, and it’s the one that turns the whole thing from a gamble into a tool.
Fourth, measurement. Some way of knowing whether the output was actually good, which then feeds back into the spec.
It’s a loop, not a line
The part that’s easy to miss: the spec isn’t supposed to be tight out of the gates. You get there by looping. Measurement feeds back into the type specification, the spec sharpens, the next pass comes back better. If you treat the recipe as a straight line and judge the first output as the verdict, you’ll conclude the tool doesn’t work. It works on the second and third lap.
Why the name matters
A handoff is “here, you finish this,” and then you walk away. A handshake is “we’re doing this together, and there’s a moment where I check the work and we agree before it moves on.” The human checkpoint isn’t a courtesy bolted onto the end or a safety net for when things go wrong. It’s structural. When I described it this way to a colleague, his reaction was that the handshake-versus-handoff line is the part worth repeating, and I think he’s right.
The teachable part
Most failed AI experiments fail on the first and third ingredients. People hand the model a vague ask with no type specification, then accept whatever comes back with no checkpoint. Add those two and a slot machine becomes a tool. Delegation is a handoff. Working with AI well, at least for now, is a handshake. The day that stops being true, the recipe changes. We’re not there yet.