Journal · The AI Architect · 2026-07-17
Read every plan before approving. Ask: do I understand it? Is it what I asked for? Could it lose anything?
Read every plan before approving. Ask: do I understand it? Is it what I asked for? Could it lose anything?
The tension isn’t whether to check the AI’s work. Everyone agrees you should do that. The tension is that checking feels like it defeats the purpose. You asked for a plan so you wouldn’t have to think as hard, and now you’re being told to think just as hard, only later, and about someone else’s reasoning instead of your own. That friction is real, and pretending it isn’t is how people quietly stop checking.
Understanding is not the same as agreement
The first question — do I understand it — gets skipped more than any other, because it feels redundant. You read the plan, it sounds coherent, you nod along. But nodding along is not the same as being able to explain it to someone else. A plan can be fluent and wrong at the same time. Fluency is cheap for a language model; it’s what the model is best at. Understanding, on the other hand, means you could redraw the plan from memory, in your own words, and have it still make sense.
Try this: after reading a plan, close it and write three sentences summarizing what it will actually do, in plain language, as if explaining it to a colleague who wasn’t in the room. If you can’t do that, you don’t understand it yet — you’ve just absorbed its tone. Go back and reread the parts where your summary got vague. That vagueness is usually pointing at the exact spot where you’d have caught a problem later, at a worse time.
Match the plan to the ask, not to the vibe
The second question — is it what I asked for — sounds obvious but hides a subtler trap. AI plans are agreeable. They tend to produce something that satisfies the spirit of your request while quietly substituting easier or more generic versions of the specifics. You ask for a plan that keeps a particular constraint intact, and you get a plan that mentions the constraint, addresses something adjacent to it, and moves on. It reads like a yes. It isn’t always a yes.
The fix is to go back to your original request — not your memory of it, the actual words — and check line by line. Did every constraint show up somewhere in the plan, addressed specifically, not just acknowledged? If you asked for three options, are there three, or did two collapse into variations of the same idea? If you asked it to preserve something, does the plan actually protect that thing, or work around it?
Try this: before reading the plan, write your request’s non-negotiables as a short checklist, three to five items. After reading the plan, check each one off only if you can point to the specific sentence that satisfies it. An item you can’t check off isn’t a paperwork problem — it’s a sign the plan solved a slightly different problem than the one you posed.
Ask what the plan quietly leaves behind
The third question is the one people skip most, because it requires imagining absence rather than reading presence. Could it lose anything. A plan can satisfy your literal request and still discard something you valued without ever naming it as a requirement — a relationship, a piece of context, a safeguard that existed for a reason you didn’t think to state because it seemed too obvious to mention.
This is where experience matters more than intelligence, artificial or otherwise. The model doesn’t know what you’d miss, because you didn’t tell it, and it has no way of guessing what wasn’t said. Your job in this step is not to evaluate what’s on the page. It’s to ask what’s not on the page that should be.
Try this: read the plan once more, specifically looking for what it removes, skips, simplifies, or postpones. Make a short list of those omissions, even the ones that seem minor or clearly intentional. Then ask, for each one, whether you’d be comfortable explaining that loss to the person most affected by it. If the answer is no, that’s not a detail. That’s a decision the plan made on your behalf, without asking.
None of these three questions takes long once you’ve practiced them. What takes time is building the habit of asking them at all, especially when the plan looks polished and your instinct is to trust the polish. Polish is not a proxy for soundness. It’s just what fluent text looks like.
This is a small piece of a larger habit — knowing when to slow down with a tool that’s built to make you feel like you can speed up. The book goes further into how to build that judgment so it doesn’t feel like extra work every time, but the three questions here are enough to start noticing where your own approvals have been running on autopilot.
Go deeper. The full method is in The AI Architect. New here? Start with the free companion pack, or explore the series.