The Monday report, written by itself

Reading time: about 10 minutes

The situation. You know how the week went by feel. If somebody asks what exactly happened in the second week of March, you do not know — and the data was there the whole time.

This case is the natural step after the sales one: there you asked once, here the question asks itself every week. And what gets automated is not the analysis — you can already do that — but remembering to do it, which is what actually fails.

What you need to hand: Having done the asking your own sales data case, because this is the same thing repeating itself. And accepting up front that the first version is manual: you automate what already works by hand, never the other way round.

Step by step

Do it by hand four Mondays in a row

Yes, four. It is not misplaced patience: until you have done it four times you do not know what you want the report to say, and automating the wrong report is wasted work that then keeps arriving every week.

Over those four times you will find there are two figures you always look at and three you asked for on day one and never read again. The final report is those two.

And you will find the format that works for you, which is nearly always the same: half a page, five lines, and one single thing flagged. A three-page Monday report does not get read on the second Monday.

Freeze the prompt that worked

After those four weeks you have a tuned prompt. That is the one you automate, word for word. And it has to be written to work without you there: no "as we discussed last week", nothing depending on a previous conversation.

You are the management assistant of a community pharmacy.
Here are ONE WEEK’s sales and the SAME WEEK last year.

Write a HALF-PAGE MAXIMUM report, with this exact structure and nothing
else:

1. ONE SENTENCE on how the week went.
2. The 2 figures: week’s revenue, and change against the same period
   last year, in euros and as a percentage.
3. THE ONE THING I SHOULD BE CONCERNED ABOUT this week. One only. If
   there is nothing notable, write "Quiet week" and stop.
4. ONE question I should be asking myself looking at this data.

RULES:
- No generic management advice.
- Every number you write must come from the files. If it is not there,
  do not write it.
- If a figure does not add up or is missing, SAY SO instead of filling
  the gap.
- Do not invent causes. If you do not know why something fell, say it
  cannot be known from this data.
- No introduction and no sign-off. Start at point 1.

Automate the boring part, not the thinking part

The chain has three links and they do not all deserve the same effort:

Exporting — the boring bit. If your software supports scheduled exports, done. If not, it is two minutes on Monday and that is fine; it is the cheapest link and the one that gets most complicated when automated.
Analysing — the frozen prompt.
Receiving — it reaching you without opening anything.

The honest shortcut, and the one to start with: automate only the reminder. An alert on Monday at 8 saying "export the week and run it through the prompt", with the prompt inside the alert itself. It is half an hour of work instead of a whole day, it works from the first Monday and it never breaks.

If you are still doing it eight weeks later, then building the full chain is worth it. And if you are not still doing it, you saved yourself the whole day — which is exactly the point.
If you do build the full chain: what to decide first

Three things, all decided before writing a line:

What leaves the pharmacy. The file sent to a model is the clean one, with no column leading back to a person. If the automatic export pulls unfiltered columns, the automation ends up sending every Monday what you would never have sent by hand. Automated things happen with nobody watching, and that is the whole difference.

What happens if it fails. If no report arrives one Monday, do you notice? A process that fails silently is worse than none, because you assume it was a quiet week. Make it tell you when it cannot run.

What it costs. One model call a week is nothing. But if you end up sending it the whole year every Monday "just in case", that adds up. Send the week and the comparison, not the history.
Read the first automatic report suspiciously

The first Monday it arrives on its own, do not read it as a report: read it as a test. Check the revenue figure against your software. If that matches, check last year’s — the one most often wrong, because it depends on the export picking the right dates.

And repeat that check one Monday a month, forever. It is the only thing that stops an automatic report turning into fiction: nothing warns you when an export starts pulling the wrong date range, and the report will keep arriving just as punctual and just as well written.

A report that is never checked stops being information and becomes a habit.

A full example, and what is wrong with it

The frozen step 2 prompt over a normal week, with the same week last year alongside. This came back.

What it replies

"1. A steady week, in line with last year and without surprises.

2. Revenue: €8,140 (−2.1% against the same week in 2025, −€174).

3. The one thing I would be concerned about: skincare sales are down 31% year on year, and it is the only category with a double-digit fall.

4. The question I would ask myself: has anything changed in the display or in who covers that section?"

And now, what is wrong with that reply:

It follows the structure and the cap, which is already a lot. Four points, half a page, no introduction and no sign-off. That is exactly what was asked and it is what keeps a weekly report being read in week twelve. The cap is not cosmetic: without it, each week it writes a little more because more text looks like more work.
But point 1 and point 3 contradict each other. "No surprises" at the top and a 31% fall below. It is not a calculation error: the summary sentence is written looking at the total and the finding is written looking at the detail, and nobody checks they agree. If you only read the first line — which is what ends up happening — you miss precisely what the report exists to tell you.
The 31% has no volume context. A 31% fall on €400 is €124 and is nothing; on €3,000 it is €930 and is the whole week. A percentage with no euro figure beside it cannot be acted on, and the prompt asked for both. Ask again with the two always together: it is a one-line correction and it changes what you do with the report.
And the question in point 4 is genuinely good. "Has anything changed in the display or the team?" is a question you can answer tomorrow and the model cannot. It invents no cause — which was the instruction — and leaves you the job of looking where it matters. That point is what makes the report worth five minutes of reading.

The correction is two sentences in the prompt: "point 1 must be consistent with point 3" and "every percentage carries its euro figure beside it". You change them once and they hold for the next fifty weeks — which is the entire advantage of freezing a prompt.

And from that comes this case’s rule: the prompt gets tuned from the first weeks’ reports, not before. The failures to fix are not visible reading it; they are visible reading four reports in a row.

And if your pharmacy is not like that

If you do not have last year’s figures. Compare against the average of the previous four weeks, which is worse and works. What does not work is comparing against nothing: a figure on its own does not say whether the week was good, and then the report becomes a number you look at and forget.

If you want the report for a team meeting. Change point 4: instead of "a question I should ask myself", ask for "one concrete thing the team can do this week". It is a different report with the same structure, and the change of audience changes what is useful — a figure nobody can act on at the counter is no use in a meeting.

If your week does not run Monday to Sunday. Say so in the prompt. In a pharmacy with weekend duty or heavy month-start activity, the standard weekly cut splits the patterns you want to see in two. A badly placed cut produces findings that belong to the calendar and not to the business.

If the report comes only to you. Then cut it to two points: the figure and the thing to worry about. Points 1 and 4 exist to show somebody else or to force you to think; if it is only for you and you are twenty weeks in, you know half the report before opening it.

When it does not work first time

I forget on Mondays and end up abandoning it.

That is why step 3 recommends automating the reminder before the chain. An 8 a.m. alert with the prompt inside costs half an hour to set up and solves 90% of the problem — which is not doing the report, it is remembering. If you are still doing it eight weeks later, then build the rest.

Every week it gives me a different format.

That is the sign the prompt is not frozen: rewrite it each Monday and each Monday something different comes out. Save it somewhere and paste it verbatim, without changing a comma. Comparability between weeks is worth more than this week’s marginal improvement.

It writes "quiet week" five weeks running.

It may be true and it is a good sign — that exit exists in the prompt for a reason. What to check is that the report does flag something when there is something: on a week you know had a drop, see whether it catches it. If it does not, the report has spent five weeks saying nothing is happening because it is not looking properly.

I am not sure it is worth the time.

Ask yourself at eight weeks with a concrete test: has there been anything you did differently because of a report? If the answer is no, drop it — and that is fine. Better to abandon a habit that does not work than keep it out of inertia and end up opening neither.

Before you use what it gave you

You did it by hand for several weeks before automating anything.
The prompt works without depending on any previous conversation.
The file leaving the pharmacy carries no personal data.
You find out if the report does not arrive one Monday.
You checked the first automatic report’s figures against your software.
You have decided when you will check them again.

It fills gaps instead of reporting them. If one Monday the export comes out empty or partial, a model will still write a coherent report. That is why the prompt tells it to say so — and why you check the figures now and then even after months of it working.

It invents the cause of a dip. "The dip is due to the end of the allergy season" sounds right and does not come from the data: it comes from you asking for a report, and a report has to explain something. The "do not invent causes" instruction is one of the most noticeable when missing.

It grows on its own without a cap. Without the half-page limit it writes a bit more each week, because more text looks like more work. By week five you stop reading it. The cap goes in the prompt and gets checked when you read it.

It compares periods of different lengths without saying so. If one of the two weeks had a public holiday and the other did not, the comparison is six days against seven and the 2% fall is actually a rise. The model will not mention it because it does not know your holidays. Always put the effective trading days of each period in the context, as in the sales case.

Now, with your own. Start with the reminder, not the chain. Set an alarm for next Monday with the prompt written inside it. If you are still doing it in two months, build the rest; if not, you have learnt something useful about yourself for five minutes’ effort.

← All practical cases Back to the school →

The Monday report, written by itself

Step by step

Do it by hand four Mondays in a row

Freeze the prompt that worked

Automate the boring part, not the thinking part

If you do build the full chain: what to decide first

Read the first automatic report suspiciously

A full example, and what is wrong with it

And if your pharmacy is not like that

When it does not work first time

Before you use what it gave you