28 APR 2026 · 8 MIN READ
The 5-prompt weekly reviewI run every Sunday.
Most weekly reviews collapse into a vague journal entry. This is the five-prompt loop I've run every Sunday for two years. Twenty-five minutes. Real decisions.
Most "weekly review" templates are bullshit. Either they are a 40-question checklist that you abandon after week three, or they are a single open-ended journal prompt that yields one paragraph of vague reflection and zero decisions.
What I want from a weekly review is simple: by 9pm Sunday, I should know what moved this week, what didn't, what to kill before Monday, and what to double down on. Twenty-five minutes maximum. No completion theatre.
This is the five-prompt loop I've run for two years. It runs in any AI that has a long enough context window — Claude is best for this because of how much context it holds, but ChatGPT works.
Before you start
The whole sequence depends on a single ingredient: a brain-dump of your week. Mine is in a Notion daily log, but the format does not matter. What matters is that you have one place where you've written down, in fragments, the things that happened and the things you noticed.
If you don't have one, the easiest start is: every evening for the next week, write 5–10 lines under three headings — "shipped today", "noticed today", "stuck on". That is enough fuel for the prompts below.
Pour the wine. Open the AI. Paste the week's notes into the first prompt.
Prompt 1 — The honest summary (3 minutes)
Here are my notes from the past 7 days [paste]. Summarise the week in 200 words. Do not be diplomatic. If the week was bad, say so. If most of the activity was busywork, say so. End with one sentence: "The week was about [X]."
The forced one-line summary at the end is the load-bearing part. It strips away the noise and asks the AI to commit to a single description of the week. If that description is "the week was about avoiding the hard thing", you now know.
I read the output once. Then I read the one-liner alone, three times. That is the diagnostic.
Prompt 2 — What actually moved (5 minutes)
From this week, list (a) the three things that materially moved the business forward, (b) the three things I spent time on that did not move the business forward, and (c) one thing I want to call "movement" but is actually just activity.
Clause (c) is the one I added after a year of running this. AI is generous by default; left alone, it will frame your busywork as progress. Asking it explicitly to call out the dressed-up activity stops the rationalisation in its tracks.
Spend a minute with each list. The (a) list usually has one entry that surprises me — something I didn't realise was the most important thing I did. The (b) list usually has at least one entry I can cut next week. The (c) line is the one I screenshot and send to a friend who'll keep me honest about it.
Prompt 3 — The decision pile (5 minutes)
Based on this week, what three decisions am I avoiding? Format: the decision, the cost of avoiding it, and the version of me that's avoiding it (which fear is doing the avoiding).
The third clause is uncomfortable. It usually outs me. "The version avoiding it is the one that doesn't want to disappoint X." "The version avoiding it is the one that wants to keep options open." Naming the version of yourself that's stalling kills 80% of the stalling.
I write the three decisions on a Post-it. Not in Notion. On paper. The act of writing them by hand is part of the prompt — it commits you to seeing them tomorrow morning.
Prompt 4 — The pre-mortem for next week (7 minutes)
It is now next Sunday. Look back on the week ahead. The week was a disappointment. Why? Be specific — name the meeting that went badly, the deliverable that slipped, the decision I deferred again.
This is the version of the pre-mortem that actually changes behaviour. The classic pre-mortem ("imagine the project failed, why?") is too abstract. Forcing the AI to imagine the specific disappointment of the specific week ahead is what surfaces the real risks.
You already wrote your calendar. The AI can see your decisions from prompt 3. Ask it to be specific. Then re-plan the week before that imagined Sunday becomes the real one.
Prompt 5 — The one rule for next week (5 minutes)
Based on everything above, give me one rule to follow for the week ahead. One rule only. Stated as: "This week, when X happens, I do Y." Make it specific enough that I can tell at the end of the week whether I followed it.
"One rule only" is the most important constraint. The week-ahead promise is small enough to actually keep. "Read more deeply this week" is rubbish. "When I'm tempted to open Slack before 10am, I write one paragraph in my draft doc instead" is a real rule.
Write it down. Pin it somewhere visible. End the review.
What this is not
This is not a productivity system. It is not a goal-setting framework. It is not a journaling practice. It is a 25-minute compression of last week into one rule for next week, run on a cadence, in the company of an AI that doesn't have to maintain your relationship with yourself.
The four ways this breaks
I have run this loop most Sundays for two years. It has broken on me four times. Each break is informative.
1. The "I'll do it tomorrow" break. You miss Sunday, plan to do it Monday morning, and Monday eats it. Then Tuesday. By Wednesday it is too late — the freshness of the week is gone. Fix: do it on Sunday, even briefly. A 10-minute version of the loop on Sunday beats a 60-minute version on Tuesday that never happens. If Sunday is impossible, do it Friday afternoon before you log off.
2. The "I have nothing to put in". You forgot to write daily notes. The brain-dump is empty. The temptation is to skip the loop. Don't — instead, use this as the prompt-1 input: "I have no notes from this week. Walk me through the seven days as I remember them and ask me one question per day to surface what mattered." It works. The lack of notes is itself a signal worth noticing.
3. The "the AI is being too positive" break. Some weeks the model leans diplomatic and the diagnostic dulls. The fix is in prompt 1 — you can sharpen it: "Be more honest than I want you to be. If the week was bad, do not soften it." If that still doesn't work, switch models for the session. Claude tends to be slightly more direct on this kind of work. ChatGPT is excellent for prompt 5 (the rule).
4. The "I followed the rule but nothing changed" break. You did the loop, kept the one rule, and the next week looked the same. Two possibilities. Either the rule was too small (it changed nothing because nothing was at stake), or the underlying problem is structural (the rule was right but the calendar was the same as last week). Diagnose this in prompt 2 of the next week's review — explicitly ask: "Did last week's rule actually move anything?" If not, the next rule needs to be bigger or the calendar itself needs to change.
The compounding bit
The output of this loop after one week is fine. After three months, it is something you would not give up. The compounding asset is the running pile of "one rule for the week" entries. After a quarter, you have 13 rules. Most of them you broke. A few of them stuck and quietly became habits. The ones that stuck are the most accurate self-portrait of where you are growing.
I keep my running list in a single Notion page titled "Rules". It is the document I read most often when I'm stuck on something — not because it has the answer, but because it reminds me which version of myself I was three months ago and what I was trying to fix.
The other compounding asset is more subtle: the AI's model of you. After three months of running the loop in the same chat, the AI has read your weeks, watched the rules, noticed which ones held and which collapsed. The diagnostics get sharper because the context is richer. This is the only "AI memory" feature that has earned its keep in my workflow — not because the tech is exotic, but because the loop is regular enough that the model has something to hold on to.
Where this fits
The full operator's version of this loop — with high-stakes-week variants, the cadence-breaks playbook (above), and a worked example from a real founder week — is queued for Bible 04 (AI Agency Accelerator), drafting now.
If you want to start using AI for ops decisions today, the free chapter from Bible 02 (Digital Publishing Empire) is the cleanest entry point. It runs you through a full publishing model end-to-end and the prompt patterns are the ones that show up across the rest of the library.
The thing that matters is doing the loop. Tonight. Before Sunday turns into Monday and the next week buries this one. Open the chat. Paste your notes. Run prompt 1. The other four will follow.