How to build an AI-assisted weekly review system

Build a Weekly Review System AI Can't Run For You

Split the review in two. Let AI handle gathering, like Granola compiling the week's meeting notes and NotebookLM surfacing themes across months of journal entries, then do the deciding yourself. David Allen's 30-to-60-minute weekly review still hinges on human reflection. The tools clear your desk; you choose where to point next week.

I treat my Friday review as the most leveraged hour of my week, and the fastest way to ruin it is to hand the whole thing to an agent. The review has two phases that feel similar but are completely different: gathering and deciding. AI is genuinely good at the first and quietly corrosive on the second, so the system I run keeps them apart on purpose.

The gathering phase is where the tools earn their keep. David Allen's Getting Things Done calls the weekly review the habit that makes the whole system trustworthy, and most of the friction is just collecting scattered inputs. So I let machines do it. Granola, which added its Recipes feature in late 2025, compiles every meeting from the week into clean summaries and pulls out the commitments I made out loud and then forgot. Todoist's AI assistant flags what slipped and breaks stalled projects into next actions. OmniFocus 4 has a dedicated review mode that surfaces every project I haven't touched. None of this requires intelligence so much as patience, which is exactly what I don't have at 4pm on a Friday.

Then there's the layer most people miss: pattern recognition across time. I keep a running journal, and once a month I point NotebookLM at the last several weeks of entries and ask, plainly, what themes keep recurring and what I keep avoiding. Because its context window is far larger than a single chat thread, it reads months at once and catches drifts I can't see from inside the week. Claude Projects does something adjacent if you load your notes as project knowledge and ask it to map reflections against your stated goals. This is the part that surprised me most as a coach: a model reading your own words back to you, organized by theme, is a strangely effective mirror. It doesn't judge, it just patterns.

But here is the line I won't cross, and it's the whole point. The deciding phase, choosing what matters most next week and what to deliberately drop, stays mine. This is where my coaching training and my builder instinct agree. John Whitmore's GROW model, the backbone of Coaching for Performance, works because the person answers their own questions about goal, reality, options, and will. The insight is generated, not retrieved. The moment I ask an AI "what should I prioritize," I've outsourced the one cognitive act the review exists to produce. The tool can lay out the options; only I carry the context of what I actually want this quarter to mean.

There's a failure mode I watched myself fall into, so I'll name it. AI makes the gathering phase so frictionless that the review degrades into a passive reading exercise. You skim a beautiful auto-generated summary, nod, and close the laptop feeling reviewed without having decided anything. That's worse than no review, because it manufactures the sensation of control without the substance. Daniel Kahneman would recognize it instantly: a fluent, effortless summary feels true and complete, and that System 1 ease is precisely what dulls the slower, harder thinking a review is supposed to force. The fix is mechanical. I make myself write three sentences by hand at the end, no AI involved: what worked, what I'm changing, and the single thing next week is actually about. If I can't write them, I haven't reviewed.

The practice of pulling back to see the larger pattern is something I dig into in the art of discovery, and the weekly review is where that habit lives at the smallest scale. So my full system is three moves. First, let your meeting and task tools gather everything automatically before you sit down, so you start with a clean desk instead of a blank page. Second, run a monthly pattern pass with NotebookLM or Claude over your own writing, treating the output as raw material, never as a verdict. Third, close every review by hand-deciding the one thing that matters and what you're consciously not doing. Build it this way and AI gives you back the twenty minutes of administrivia while leaving you the ten minutes of judgment that were always the real work. Hand it the whole hour and you'll have a tidy archive of a life you stopped steering.