How AI finally made the weekly review stick

In 2001, David Allen published Getting Things Done. Buried about two-thirds of the way through the book, past the inboxes and the contexts and the 2-minute rule, is the single idea that everyone agrees is the most important thing in it: the weekly review.

The concept is simple. Once a week, you sit down for an hour. You empty your inbox. You go through your projects list and decide what's still active. You look at your calendar for the past week and the week ahead. You reread your goals. And then — the part almost nobody actually does — you ask yourself whether the last week moved you toward the life you said you wanted.

Allen was right. The weekly review is the heart of the whole system. It's the thing that turns a list of tasks into a life you're actually building.

And yet, in twenty-five years of GTD preaching, the weekly review remains the single thing people skip most. Surveys of GTD practitioners consistently find that fewer than 20% of self-identified GTD users actually do the weekly review on a weekly basis. The rest of us mean to. We block the time on Sundays. We feel vaguely guilty all week. Then Sunday arrives and we watch football instead.

This is not because we're lazy. It's because the weekly review, as Allen conceived it in 2001, has a friction problem that no amount of discipline can fix. Until now.

The friction problem

Try to imagine what an honest weekly review looks like. Not the five-minute version — the real one, the one Allen described. You'd need to:

1. Pull up every task you completed in the last seven days
2. Pull up every task you didn't complete
3. Pull up your calendar and count hours by category
4. Remember how you felt each day, or at least a few key moments
5. Reread your stated goals
6. Compare (1)-(4) against (5)
7. Notice the gaps honestly
8. Decide what to do about them

Steps 1 and 2 take maybe ten minutes if your app is well-organized. Step 3 takes another ten minutes of staring at your calendar. Step 4 is nearly impossible — most people can't reconstruct yesterday, let alone six days ago. Step 5 is quick. Steps 6 through 8 are the actual valuable part of the exercise, and they require the first five steps to be done well enough that you have something to compare against.

So the weekly review, done properly, takes an hour on a Sunday afternoon and is emotionally draining and produces an output that often makes you feel bad about yourself. Of course people skip it.

The solution in 2001 was discipline: force yourself to do it anyway, because it's the most valuable hour of your week. This worked for maybe 10% of people who read the book. For everyone else, it became a source of low-grade guilt for twenty years.

What AI actually changes

Here's the thing about steps 1 through 5 in the weekly review process: they're all retrieval and summarization. They don't require judgment, creativity, or wisdom. They require patience and completeness. They're exactly the kind of task that's tedious for humans and trivial for a modern LLM.

Large language models circa 2026 — specifically the tool-using, long-context variants like Claude — are genuinely good at three things relevant to weekly reviews:

• Reading a week's worth of structured data (calendar events, completed todos, voice transcripts, transactions) and producing a concise summary in natural language.
• Comparing that summary to a stated set of priorities and noticing when they don't match.
• Writing a short, specific, kind recommendation for what to do next week to close the gap, in a tone that doesn't make you want to throw your phone.

These are exactly the three things the human in a weekly review is supposed to do. And the AI does them in roughly two seconds.

This is not a minor improvement. This is the difference between "the hour-long review I keep meaning to do" and "the notification that shows up Monday morning while I'm making coffee." It's the same shift that happened to email search when Gmail launched — not a faster way to do the thing, a categorically different relationship with the thing.

What goals. actually does on Monday morning

Here's what happens inside goals. every week. I'll be specific because most "AI productivity" claims are vague enough to be meaningless.

1. Aggregate the week's signals

Every Monday at 8am (or whenever you've scheduled it), a server-side function runs. It reads:

• Every todo you completed or created in the last seven days, tagged by which goal it linked to
• Every voice reflection you recorded, with its on-device transcript
• Every calendar event (if you've connected Apple Calendar), bucketed by goal category using a keyword classifier plus Claude's own categorization
• Daily health aggregates (if you've connected Apple Health) — steps, active minutes, sleep hours, workouts. These are derived on-device; raw health samples never leave your phone
• Any spending transactions extracted from credit-card statement uploads, categorized against your goals

2. Compute a per-signal alignment score

Each signal gets a 0–100 score that answers "did this area of the week match your stated priorities?":

• Time allocation score: the overlap coefficient between your actual calendar hours per goal category and your intended allocation based on goal priorities
• Sentiment score: the average sentiment and energy level from your week's voice reflections
• Spending score: the ratio of transactions that support your goals vs. contradict them
• Health score: percentage of days that qualify as "healthy" (7k+ steps, or 20+ active minutes, or a workout, and sleep >= 6.5 hours if reported)
• Action completion score: the percentage of last week's suggested actions and your own this-week todos that actually got checked off

These get rolled up into a single overall alignment score via a weighted average. Missing signals drop out of the weighting cleanly — if you only use todos and voice, you still get a meaningful score.

3. Ask Claude for a two-paragraph briefing

The server packages your goals, your raw time-allocation totals, and your per-signal scores into a JSON payload and sends it to Claude Haiku with a careful prompt. The prompt basically says: "You're writing a short weekly briefing for the user. Don't recite the scores. Notice the biggest gap between their stated priorities and their actual behavior. Write a two-sentence recap in plain language and one or two concrete micro-actions they could take this week to close the gap. Be honest but not harsh."

Claude returns a JSON blob with the summary and 1–2 actions. The actions are specific ("Block two deep-work mornings on Tues/Thurs") not vague ("Focus more on deep work"). If an action relates to a specific goal, Claude includes a link back to that goal row so the UI can surface it contextually.

4. Push the briefing to your phone

The briefing gets stored in the database as a check_in_cards row, and a local notification fires to tell you it's ready. You open the app, read a short honest paragraph that took Claude ~1 second to write, and either check off the suggested actions or dismiss them.

Total time from your end: about 30 seconds, on a Monday morning when you were already drinking coffee.

The interesting second-order effect

Here's what surprised me when we started using this internally. The value isn't really the briefing itself — the briefing is often something I could have figured out by thinking about my week for thirty seconds. The value is that the briefing happens reliably, every week, without me having to decide to do it.

Most self-improvement loops fail at the decide-to-do-it stage. You know you should meditate. You know you should journal. You know you should do the weekly review. You just don't, because the moment of deciding is a moment of friction, and friction wins. The AI weekly review removes the friction of deciding entirely. It's going to happen whether you engage with it or not. You might as well engage.

And once the review is happening reliably, a funny thing starts to happen: you start adjusting your behavior in advance because you know the Monday briefing is going to read your week accurately. You start scheduling the fitness block on Wednesday because you know the briefing will notice you didn't exercise last week. You start making the hard call on the project that isn't serving a goal because you know Claude is going to ask about it.

This is, in the technical literature, called accountability via legible feedback, and it's the same reason Strava makes you run more even when nobody else is looking at your workouts.

The part that isn't AI

The briefing is written by Claude. But the scoring isn't, and the goal-to-todo linking isn't, and the voice-reflection sentiment analysis is. The four-signal framework (todos, voice, calendar + health + spending) isn't a Claude invention — it's a deliberate design choice about what's worth measuring and what's noise. The weekly cadence isn't an AI thing — it's David Allen from 2001.

What AI adds is the synthesis step. Taking all those inputs and producing a short, honest, kind paragraph used to take a disciplined human an hour. Now it takes a machine a second. That's the change.

The humans didn't become unnecessary. The discipline became unnecessary. That's better, because discipline is the rarest resource most of us have.