# Paste this first (then your data)

Send this message before you paste any spreadsheet export into ChatGPT, Claude or a local model. It makes the AI read the columns back to you, refuse to count by eye, and tell you what it cannot see. From Distru's No Bullshit AI Course, lesson 03-02. CC0 1.0.

Test status, 2026-09-17: the four lines map one to one onto the failure list in course lesson 01-02 (misread column, personal data, arithmetic over many rows, invented values). We have not benchmarked it across models. Run it on your own export first and read what comes back before you trust it.

Replace the parts in `[brackets]`. Then paste your header row and rows below it in the same message, or in the next one.

---

```
I run [a dispensary / a cultivation / a distributor] in [state]. I am about to paste a spreadsheet export from [Distru / Metrc / my POS / QuickBooks]. It has [N] rows and these columns: [paste header row].

Before you answer any question about it:
1. Describe each column in one line: what it seems to hold, its unit if any, and how many blanks you see. If you cannot tell what a column is, say so; do not guess.
2. Tell me which columns look like they could identify a person. I will remove them before we go further.
3. For any question that needs counting, adding, averaging or sorting more than about 20 rows, do not compute the number yourself. Write the spreadsheet formula or a short [Python / JavaScript] script that computes it, and tell me in two sentences how to run it and how to check that it worked.
4. Quote the exact rows you used for any claim you make. If a value you need is missing, write MISSING instead of guessing.

Here are the first rows:
[paste header + rows]
```

---

## Why each line is there

- **Line 1** catches the most common failure: the AI misreads a column (treats `quantity` as eaches when it is grams) and everything after is wrong.
- **Line 2** is your redaction backstop, not your redaction. Remove personal data before you paste; this only catches what you missed.
- **Line 3** is the rule that matters. Language models are poor at arithmetic over many rows and sound confident while wrong. A formula is checkable. A number is not.
- **Line 4** gives you something to verify. Open the rows it quoted and look.

Pair it with `csv-head.py` in this folder, which prints the header, row count and a token estimate so you can fill in `[N]` and `[paste header row]` without opening Excel.