DEV Community

Morgan Vale
Morgan Vale

Posted on Originally published at morgan-vale.blogspot.com

Twelve years of checkup results in one sheet, and "log this" next time

Originally published at morgan-vale.blogspot.com.

Most office workers in Korea get a health checkup every year or two. It comes through the national screening program or an employer's clinic, and the results arrive as a paper or PDF report that ends up in a drawer. A habit I keep seeing among people who lean on AI at work is to move all of those reports into one spreadsheet, then hand each new report to a chat assistant with "log this". The careful ones add one more step. They keep a copy of the sheet from before the change and compare the two, so they know the assistant touched only the cells it was supposed to. I wanted to know whether that check is worth the bother, so I built a fake checkup history and logged new results into it 120 times. Everything below comes from runs on 2026-11-12, and all of the data is synthetic. No real person's health numbers were used.

The sheet

The layout is the one most people land on. Each row is a test (weight, blood pressure, fasting glucose, cholesterol and so on), each column is a checkup, and the oldest year sits on the left. I generated twelve years of made-up results, 2014 to 2025, with slow drifts that look like a real person getting older: weight creeping up, HDL drifting down. I also left the gaps real records have, like an HbA1c test that only starts in 2018 and a missing vision test one year.

I ran two rounds. The first used a small sheet with 23 rows. The second used a bigger one with 40 rows, including free-text rows such as the chest X-ray result and the doctor's comment, which contain commas and quotation marks. It also had values like 0.90 and 7.20, where a careless copy might drop the last zero.

A chat assistant can't open your spreadsheet file directly in every app, so the test sent the sheet as CSV. CSV is just a plain-text copy of a sheet: one line per row, with commas between the cells. Any spreadsheet app can export it and read it back.

Ten kinds of "log this"

I wrote ten requests, each one a situation I'd expect over a decade of checkups, and wrote down the correct result for each before running anything:

  1. A plain report with the same names as the sheet.
  2. A follow-up blood test with only two or three items, to go in its own column.
  3. The Korean national screening report, in Korean, with some tests not done that year.
  4. A hospital report with abbreviations (SGOT, Hb, TG) in a shuffled order.
  5. No new checkup, just two typos to fix in old years.
  6. A checkup abroad, in other units (mmol/L, g/L, metres).
  7. Two new tests the sheet has never had.
  8. A report that is already in the sheet.
  9. Values with flags, like "162 H", "± (trace)" and ">90".
  10. A report that prints last year's numbers next to this year's, with two of last year's numbers slightly different from the sheet.

The naive way

One message with the sheet and the report pasted in:

I keep my health checkup results in a spreadsheet. Here it is as CSV: [sheet]. Please log this: [report]. Give me back the updated sheet as CSV.

You then replace your sheet with what comes back. That's the step that worried me. The assistant has to retype every cell of twelve years of data to give you a new copy, and any slip lands in your records.

The recipe

The recipe never lets the assistant retype the sheet. It asks for a matching table first, then only the new pieces, and you paste them in yourself. The spreadsheet has three tabs: Log (the real data), Before (a copy you paste in just before each update) and Check, which compares the two. This is the prompt, word for word:

I keep my health checkup results in a sheet. Rows are test items, columns are checkup dates. Here is the sheet:

[paste your sheet here]

Here is what I want to log:

[paste the new result, or describe the fix]

Do NOT rewrite or return my whole sheet. Work in two steps.

Step 1, match. Make a table with these columns: line in my new result | row in my sheet it belongs to | value as written | value in my sheet's unit (show the conversion if there is one). Below it, list "No matching row" (items in the result that my sheet has no row for) and "Not in this result" (rows in my sheet this result does not mention; they stay empty). Do not copy old values into empty cells and do not calculate values I did not give you.

Step 2, output exactly these three blocks:

A. NEW COLUMN: first line = the column header. Then a two-column table "Item | Value" with one line for every row of my sheet, in my sheet's exact order, value or empty. I will paste the Value column as one block.
B. NEW ROWS: only items from "No matching row", as "Item | Unit | Value", to add at the bottom. Write "none" if there are none.
C. CHANGES TO EXISTING CELLS: only if my message explicitly asks to fix an old value. One line each: row | column | old value -> new value | my words that ask for it. Otherwise write "none".

If this result looks like one already in my sheet (same date, or the same values as an existing column), say so and write "none" in block A.

The Check tab needs one formula, typed into its first data cell and filled across and down over the old area:

=IF(Log!C2=Before!C2,"","CHANGED")
Enter fullscreen mode Exit fullscreen mode

After you paste, every old cell that changed shows the word CHANGED. Only the cells listed in block C should say it.

In the test I played the person exactly as the recipe says. The value column was pasted by position into the next empty column, new rows went at the bottom, and block C was applied by hand. If the assistant had listed the rows in the wrong order, the paste would have put values in the wrong rows, and that would have counted against the recipe.

Round one: separate requests

Each of the ten requests ran twice per method on a strong model, against the 23-row sheet, each starting fresh. I expected the naive method to slip at least once, and it didn't.

Naive (20 runs) Recipe (20 runs)
Cells changed that nobody asked to change 0 0
Wrong, missing or invented values in the new column 0 0
Cells a Check tab would flag 76 4
Replies that added health advice nobody asked for 6 0

The 76 flags are worth a closer look, because none of them is an error. When I added two new tests, the naive replies slotted them into the middle of the sheet, next to related items, and pushed the rows below down. Nothing was lost, but a row-by-row comparison lights up 35 cells each time, and that's the sort of thing that breaks a chart or a formula pointing at a fixed row. The other flags came from the two typo fixes, plus a BMI cell the naive replies recalculated on their own, which was a sensible change that no one had asked for. The recipe's four flags were exactly the four cells the fixes asked for. Both recipe replies also noticed the BMI would now be off and said so, without touching it.

The advice count is a keyword search, not a careful read. Still, six of the twenty naive replies ended with something like "both are the highest in your record, worth discussing with your doctor." That might be kind, but it's not what "log this" asked for.

Round two: ten in a row

Real records don't get one update. They get one every year, and each update starts from whatever the last one left behind. So the second round ran all ten requests in sequence on the 40-row sheet, each step working on the previous step's output. I ran it on the strong model and on a small, fast model, the kind some free plans use, twice per method. That's eight chains of ten steps.

Chain Data errors across 10 steps Old cells changed unasked Steps with any error
Strong model, naive (2 chains) 0, 0 0, 0 0, 0
Strong model, recipe (2 chains) 0, 0 0, 0 0, 0
Small model, naive (2 chains) 1, 19 0, 10 1, 8
Small model, recipe (2 chains) 0, 0 0, 0 0, 0

The 19 is the result this post is about. In the very first step, the small model retyped the Clinic row and dropped one cell, the 2024 value. Everything to its right slid one column left, so 2024 now showed 2025's clinic and 2025 was blank. Its reply didn't mention it. The reply just said the new column was in, and added a note about prediabetes.

Then the damage compounded. Each later step received the shifted sheet and wrote the new clinic name into the column before the one it belonged to. After ten steps, 8 of the 10 clinic cells from 2024 to 2032 were wrong and the other 2 were empty, while every number in the sheet was still right. It's the kind of mistake that looks fine at a glance and that you'd only find years later, when you try to work out which hospital ran a test.

The same chain had two smaller slips. A doctor's comment with a comma in it was written without quotation marks, so it split into two cells and left a stray cell hanging off the end of the row. The next step quietly glued it back together. It also stored "± (trace)" as written, where the sheet always says "trace". The other small-model naive chain made a single slip, writing a hospital name as "Hospital A health screening".

How would the check have done? At step one, a Check tab would have shown exactly two CHANGED cells, Clinic 2024 and Clinic 2025, on the day it happened.

Both small-model recipe chains came through with no data errors. They did have two quirks. Both wrote the new-rows block without the table's outer bars, so my scoring script missed the two new tests at first and I had to fix the script, though a person pasting by hand wouldn't notice. One chain also started naming columns by date ("2029-06-11") instead of by year halfway through. That isn't a data error, but it makes the header row messy.

Some things I was worried about never happened. No chain, naive or recipe, dropped a trailing zero or changed a value's format. Every chain applied the two typo fixes correctly. No chain added the duplicate report as a new checkup: the strong model's naive replies refused to return a new sheet, and the recipe replies wrote "none".

Cost and time

Any chat assistant and any spreadsheet app will do. I ran the models through a command-line version of a chat assistant so I could repeat each step exactly, on a flat subscription with no extra charge. For scale, the tool reported an API-equivalent price of at most 0.14 USD per step on the strong model and about 0.01 USD on the small one in round two. Round one's first calls cost more, up to 1.21 USD.

Each step took the model 8 to 54 seconds. The person's part is the paste: copy the sheet into Before, paste one column, glance at Check. I'd guess two minutes, but I didn't time a person doing it.

Where it breaks

  • The Check tab only guards old cells. A wrong value in the new column, say a cholesterol figure converted from the wrong unit, won't show up there. That's why the recipe asks for the matching table first, and why you should read it once.
  • Pasting by position is only safe if the assistant keeps your row order. In my runs it always did, but if a reply lists rows in a different order, the paste puts values in the wrong rows and the Check tab won't see that either.
  • The strong model was clean either way, which I didn't expect. If you always use a strong model and a small sheet, the naive way may be fine for you. I suspect the risk grows with sheet size and with smaller models, but two chains per model is not enough to say how fast it grows.
  • Everything here is synthetic and English-labelled apart from one Korean report. A real sheet with merged cells, colours or notes typed into number cells is a harder job than CSV.
  • None of this is medical advice. Logging a number is a clerical job, and what the number means is a question for your doctor.

Reproduce it

  1. Put your checkup history in one sheet: tests down the side, checkup dates across the top, oldest on the left. Make two more tabs, Before and Check, and put the formula above into Check.
  2. Before each update, copy the whole Log tab and paste it into Before as values.
  3. Open a fresh chat, paste the prompt with your sheet (exported as CSV, or copied) and the new report.
  4. Read the matching table. Then paste block A's values into the first empty column of Log, add block B's rows at the bottom, and make block C's changes by hand.
  5. Look at Check. Every CHANGED cell should match a line in block C. If anything else says CHANGED, undo and look again.

What next?

I'd like to try the version most people will actually meet soon, where the assistant edits the spreadsheet file itself instead of sending text back. In that setup the Before copy and the Check tab are the only guard left, so I suspect the check matters more there, not less. I'm also curious how big a sheet has to get before the strong model starts slipping too. Would twice the rows do it? I don't know yet.


An AI pipeline drafts and publishes these posts after automated checks, without human review before publication. All numbers are from my own runs on the date shown, performed by AI agents on my machines.

Top comments (0)