Most weekly reviews fail for a boring reason: the notes pile up, and rereading them takes longer than the insight is worth. So the review gets skipped, or it gets replaced by a vague feeling (“this week was rough”) that changes nothing.
A model is genuinely useful here — not because it can tell you what your week meant, but because it can scan forty scattered notes in seconds and group the ones that repeat. The failure mode is asking it the wrong question. “What should I change this week?” produces a fluent, generic answer that sounds like advice and is really just pattern-matched self-help language with your nouns swapped in.
The better question separates two jobs that people usually blur together: finding repetition, and deciding what the repetition means. A model can do the first well. Only you can do the second.
The reframe: evidence first, judgement second
Instead of asking for conclusions, ask for candidates with receipts.
Weak prompt:
Here are my notes from this week. What patterns do you see, and what should I do differently?
This invites the model to produce a confident-sounding narrative arc — “you seem to be struggling with focus and boundaries” — regardless of whether the notes actually support it. Fluent framing is not evidence.
Stronger prompt:
Below are my notes from Mon–Sun, pasted as-is with timestamps.
Do not give advice or conclusions.
1. List candidate patterns — anything that appears 3+ times this week.
2. For each candidate, quote the exact lines that support it, with their dates.
3. Separately, list "open loops": things mentioned as unresolved, planned, or waiting on someone, that never got a follow-up note.
4. If a pattern only has 1–2 supporting lines, put it in a "weak signal, not enough evidence" section instead of the main list.
[paste notes]
The second version does not ask the model to interpret your week. It asks it to sort and cite. That is a task models are actually good at — and one you can check in under a minute per item, because every claim comes with the line it is based on.
Worked example (illustrative)
To show the shape of a real output, here is an illustrative week — not a real user’s notes, built to demonstrate the format:
Prompt: [as above, with five days of scattered notes: three mentions of “pushed the deck to tomorrow again,” two mentions of skipping lunch to catch a call, one calendar note about a dentist appointment moved twice.]
Model output (abridged):
Candidate pattern 1 — deadline slippage on one deliverable. Evidence: Mon “pushing deck to tomorrow,” Wed “deck still not done, moving to Fri,” Thu “deck done, later than planned.” 3 quotes, same artifact, escalating language.
Candidate pattern 2 — lunch skipped for calls. Evidence: Tue “skipped lunch, call ran long,” Thu “another lunch call, ate at desk.” Only 2 quotes — flagged as weak signal, not enough evidence for a main pattern yet.
Open loop: dentist appointment moved twice (Mon, Wed), no note of a new date or a decision to cancel.
Notice what the model did not do: it did not tell you the deadline slippage means you are bad at estimating, and it did not tell you to “set better boundaries” around lunch. It sorted and cited. The interpretation — whether the deadline slippage is about the deliverable, the estimate, or something upstream — is your job, because you have context the notes do not fully capture.
Models are fluent enough to describe a two-instance blip as “a clear pattern of overcommitment.” Set a minimum evidence bar before you accept anything — three separate, dated mentions is a reasonable default — and read every quoted line yourself before deciding a theme is real. A pattern with one strong quote and two loosely related ones is not the same as three matching quotes.
The four-step workflow
- Collect. Gather the week’s notes into one document — a notes app export, meeting notes, a running Slack-to-self channel, whatever you already use. Do not reconstruct memory; use what you actually wrote at the time.
- Extract candidates with evidence. Run the prompt above. Require quotes and dates for every candidate; require a separate “weak signal” bucket for anything under your evidence threshold.
- Verify and decide. Read the quotes. Reject anything where the quotes do not actually support the label the model gave it — this happens more than you would expect, especially when notes are terse. A common failure: three quotes that share a topic (“the deck”) but do not actually share a pattern — one is about scope, one is about a scheduling conflict, one is about someone else missing a deadline. Grouped together they look like a trend; read individually, they are three unrelated events wearing the same label. Accept only what survives a plain, line-by-line read.
- Choose one adjustment. From the accepted patterns, pick exactly one thing to change next week. Not three. One. The model can propose candidate adjustments if you ask, but the choice is yours — you know which one is actually feasible given everything else going on.
Do not paste confidential work information, health details, relationship specifics, or anything about a child into a consumer AI tool without checking your organization’s policy first. Redact names, client identifiers, and anything you would not want retained by a third party before you paste. See workplace data hygiene for a fuller checklist.
Validation and fallback
Some weeks produce nothing worth acting on — thin notes, an unusually quiet week, or a week where the real story is one big event that does not need pattern analysis. That is a valid outcome. Do not force a narrative because the ritual expects one.
A simple validation rule: if no candidate survives your evidence threshold, skip the “choose an adjustment” step entirely and just carry the open loops forward. A weekly review that occasionally produces “nothing conclusive this week” is more trustworthy than one that always finds something, because always finding something is a sign the model — or you — is pattern-matching on noise.
When to stop reviewing and talk to someone
A repeated pattern about your own work habits is a reasonable thing to adjust alone. A repeated pattern about a relationship conflict, a health symptom, or a work situation involving another person’s conduct is not something a weekly-review prompt should resolve. If the notes keep surfacing the same interpersonal or health concern, the next step is a conversation with the person involved, a manager, or a clinician — not a more detailed prompt. Emotional clarity before you respond and decision hygiene under stress cover what to do when a pattern turns out to be emotionally loaded rather than merely logistical.
Rolling weekly reviews into a monthly pattern
A single week rarely has enough evidence to justify a big change. Real patterns usually show up across four or five weekly reviews, not one. Keep the accepted patterns and adjustments from each week in one running document, and once a month, run a second pass over just that summary — not the raw notes again.
Below are my last four weekly-review summaries, each with its accepted patterns
and the adjustment chosen that week.
1. List anything that recurred across 2+ weeks, with the weeks it appeared in.
2. List adjustments I chose but did not actually follow through on — quote the
weeks where the note mentions the adjustment was dropped, forgotten, or
only partly done.
3. Do not propose a new adjustment. Just show me the recurrence and the
follow-through gap.
[paste the four summaries]
This is where the discipline pays off. A single week’s “I skipped lunch twice for calls” is a weak signal. The same line showing up in three of the last four weekly summaries is a real pattern — and one you now have dated, quoted evidence for, rather than a vague sense that “this keeps happening.” It also surfaces a different, often more useful kind of pattern: adjustments you keep choosing and keep not doing, which is itself worth naming honestly before choosing a fifth adjustment that will meet the same fate.
Make it reusable
Once the shape works, save the prompt as a reusable instruction in whatever tool you use — a Custom GPT, a Claude Project, or a plain saved snippet — so you are not re-explaining the evidence rule every week. Custom instructions and memory covers how to do that without letting the tool silently retain more than you intended.
The companion template below gives you the four-step structure as a fill-in-the-blanks sheet: candidate patterns with an evidence column, an open-loops table, and a single adjustment field with a review date. Print it, or keep it as a running document you update each week.
For decisions that come out of the review — not just noticing a pattern, but deciding what to do about it — the four-step framing in AI for better decisions is the natural next stop: frame, generate, stress-test, decide.
The part that stays yours
A model extracting three dated quotes about deadline slippage is doing pattern recognition on text you already wrote. It has no idea whether the slippage matters, whether it is about you or about an unrealistic timeline someone else set, or whether fixing it is worth the cost to something else in your week. That judgement was always going to be yours. The value of the workflow is that it gets you to the judgement faster, with less rereading and less risk of mistaking a compelling story for an actual pattern.
