Journal · The AI Architect · 2026-07-10
Describe the destination, not the route — say what "done" looks like and let your AI find the way there.
Describe the destination, not the route — say what “done” looks like and let your AI find the way there.
The real tension isn’t about trusting AI enough to let go of the wheel. It’s that most of us don’t actually know what “done” looks like until we start describing it — and the act of describing it forces a kind of clarity we usually skip. We say “write me a summary” or “clean up this code” and then feel vaguely disappointed by what comes back, without noticing that we never said what the summary was for, or what “clean” meant to us. The AI didn’t fail. We handed it fog and asked for a photograph.
Specify the outcome, not the steps
When you give an AI a sequence of steps, you’re doing two jobs at once: deciding what should happen, and deciding how it should happen. Most people are only qualified — or interested — in the first job. But because instructions feel more concrete than outcomes, we default to writing instructions, even bad ones, because at least they feel like doing something.
Try this: next time you’re about to type a numbered list of steps, stop after the first sentence and ask, “What would make me say yes, this is right, without me touching it again?” Write that sentence down. Then hand the AI that sentence instead of your list. If it’s a report, don’t say “start with an executive summary, then three sections, then a conclusion.” Say “a reader with five minutes should understand the decision and the risk in it.” Let the AI propose the structure. You review the structure against your sentence, not against a script you wrote in advance.
This is harder than it sounds, because it requires you to actually know your own criteria. Often you don’t — you just have a feeling. The discipline of writing the destination sentence is what turns the feeling into something you can check against.
Build in the checkpoints, not the checklist
Describing the destination doesn’t mean disappearing until the end and hoping for the best. It means moving your attention from the route to the checkpoints along it — the two or three moments where you actually need to see the work and say “yes, keep going” or “no, different direction.”
If you skip this and only judge the final output, you end up re-litigating the whole thing when it’s wrong, because you never said anything sooner. If you micromanage every step, you’re back to writing the route. The middle path is picking, in advance, the one or two points where a wrong turn would be expensive to unwind — and checking in only there.
Try this: before you send a request, write one line — “show me the outline before you write the full draft” or “check with me before you touch the database schema.” That’s your checkpoint. Everything before and after it, the AI figures out on its own. You’re not withdrawing from the process; you’re deciding where your judgment actually matters and being quiet everywhere else.
Say what would make it wrong
Destinations are easier to describe by their edges than by their center. It’s often simpler to say what failure looks like than what success looks like — “don’t make it longer than a page,” “don’t assume the reader knows the jargon,” “don’t touch the config file.” These constraints do a surprising amount of work, because they narrow the space of acceptable answers without you having to map the whole territory yourself.
This matters especially with anything that touches real consequences — a document going to a client, code going into production, a message going to a team. You don’t need to specify the perfect path. You need to rule out the paths that would actually hurt you. An AI with three clear boundaries and one clear target will usually find a reasonable way through, even in unfamiliar territory. An AI with a long list of steps and no stated boundaries will follow the steps precisely into a wall.
Try this: for your next request, add one sentence that starts with “avoid” or “don’t.” Notice how much lighter the rest of your instructions can be once that boundary is in place.
None of this is really about AI. It’s about the difference between managing tasks and managing outcomes, which is a skill worth having regardless of who or what is doing the work. Getting comfortable naming the destination — and trusting someone or something else to find a way there — changes how you delegate to people too. The book goes further into how to write these destination statements well, how to calibrate checkpoints for different kinds of risk, and what to do when the AI’s route surprises you in a good way. But the place to start is small: one sentence, before your next request, that says what done actually looks like.
Go deeper. The full method is in The AI Architect. New here? Start with the free companion pack, or explore the series.