# Analyse a Small CSV and Report Three Findings With Caveats

> Paste a small CSV and get three defensible findings with the numbers behind them, how data issues were handled, and what the data can't show.

- **Author:** [Wei Zhang (@wei_zhang)](https://promptabide.com/wei_zhang)
- **Tested on:** Claude · Opus 5.5
- **You fill in:** `csv`, `context`, `question`
- **Published:** 2026-08-27
- **Updated:** 2026-09-24
- **Tags:** `data-analysis`, `research`, `email-marketing`, `spreadsheets`
- **Keywords:** analyse csv data with ai, newsletter open rate analysis, best day to send newsletter data, find insights in small dataset, data analysis prompt with caveats
- **Views:** 2390
- **Likes:** 26

**Best for:** Creators, marketers and small-business owners with a small export (newsletter, sales, ads) who want a careful read, not a dashboard.

## Prompt

```
Analyse this dataset. I want three findings I could defend in a meeting, not a list of every number.

Data (CSV): {{csv}}
What the data is and how it was collected: {{context}}
The question I care about: {{question}}

Steps:
1. Describe the data in 2 lines: rows, columns, time span, and any quality issues (missing values, outliers, inconsistent units). Say how you handled each.
2. Show the calculations you rely on in a small table so I can check them by hand.
3. Give exactly 3 findings, ranked by relevance to my question. Each has a one-sentence finding, the number behind it, and a confidence level (high / medium / low) with a reason.
4. Caveats: what this data cannot show (sample size, confounders, selection effects), specific to this dataset.
5. The one additional piece of data that would most change or confirm finding 1.

Do the arithmetic carefully and double-check totals before writing. Don't claim causation unless the data design supports it.
```

## Variables

- `{{csv}}` — The data as CSV with a header row; keep it small (under about 50 rows) (e.g. issue,date,day,subject_style,recipients,opens,clicks...)
- `{{context}}` — Where the data came from and anything unusual about it (e.g. Newsletter platform export; issue 11 had a free template giveaway)
- `{{question}}` — The decision or question you want the data to inform (e.g. Should I keep sending on Tuesdays or switch to Saturdays?)

## Example input

```
csv:
issue,date,day,subject_style,recipients,opens,clicks
1,2026-06-02,Tue,how-to,4210,1850,212
2,2026-06-09,Tue,list,4290,1802,187
3,2026-06-13,Sat,question,4330,2165,260
4,2026-06-16,Tue,news,4380,1621,131
5,2026-06-20,Sat,how-to,4440,2264,301
6,2026-06-23,Tue,how-to,4490,1931,246
7,2026-06-30,Tue,question,4530,,198
8,2026-07-04,Sat,list,4600,2208,239
9,2026-07-07,Tue,how-to,4660,1957,233
10,2026-07-11,Sat,news,4700,2021,150
11,2026-07-14,Tue,list,4750,1900,1420
12,2026-07-18,Sat,how-to,4810,2405,318
13,2026-07-21,Tue,question,4850,1989,228
14,2026-07-25,Sat,list,4900,2352,245
context: Export from my newsletter platform for a weekly design-tips newsletter. I started testing Saturday sends on 13 June. Issue 11 included a free template giveaway. Issue 7's open data is missing because of a tracking outage.
question: Should I keep sending on Tuesdays or switch to Saturdays, and does the subject line style matter?
```

## Output

*Produced by Claude · Opus 5.5.*

```
**1. The data**
There are 14 issues (8 on Tuesday, 6 on Saturday) from 2 June to 25 July 2026, with recipients growing from 4,210 to 4,900. Issue 7 has no open data, so I left it out of open rates but kept its clicks. Issue 11's clicks (1,420, a 29.9% click rate against 3-7% for the others) come from the giveaway, so I left it out of click comparisons.

**2. Calculations**

| Group | Issues | Opens / recipients | Open rate | Clicks / recipients | Click rate |
|---|---|---|---|---|---|
| Tuesday | 6 opens / 7 clicks* | 11,150 / 26,880 | 41.5% | 1,435 / 31,410 | 4.57% |
| Saturday | 6 | 13,415 / 27,780 | 48.3% | 1,513 / 27,780 | 5.45% |
| How-to | 5 | 10,407 / 22,610 | 46.0% | 1,310 / 22,610 | 5.79% |
| News | 2 | 3,642 / 9,080 | 40.1% | 281 / 9,080 | 3.09% |

*Tuesday opens exclude issues 7 and 11 for missing opens and the giveaway; clicks exclude issue 11 only.

Open rate by style, Tuesday vs Saturday: how-to 42.9% vs 50.5%; list 42.0% vs 48.0%; question 41.0% vs 50.0%; news 37.0% vs 43.0%.

**3. Findings**
1. **Saturday sends get opened more:** 48.3% against 41.5%, and Saturday wins in all four subject styles, by 6-9 points. *Confidence: medium.* It's consistent, but there are only 6 Saturday issues.
2. **The Saturday click advantage comes from opens, not from more engaged readers.** Clicks per open are about the same: 11.3% on Saturday and 11.1% on Tuesday. *Confidence: medium.*
3. **"News" subject lines do worst,** with a 3.09% click rate against 4.9-5.8% for other styles. Clicks per open are also lower (7-8% against 10-13%). *Confidence: low.* There are only 2 news issues.

**4. Caveats**
- Saturdays were only tested from 13 June, but Tuesday issues from that date onwards still open at about 41%, so list growth doesn't explain the gap.
- Open rates are inflated by email apps that pre-load images, so treat them as relative, not absolute.
- Day and style weren't assigned at random. Any pattern in which styles you chose for which day could bias the comparison.
- 14 issues can't separate a real day effect from a few good topics.

**5. What would confirm finding 1**
Send the same issue on both days to a random 50/50 split of the list for 4 weeks. Most newsletter platforms support A/B sends.
```

## Why it works

**"Say how you handled each"** quality issue forces decisions about outliers and missing values into the open, instead of an average quietly skewed by a giveaway. The **calculation table** lets you check the arithmetic by hand, which matters because models can slip on sums. **Exactly 3 findings with confidence levels** makes the model prioritise instead of listing everything. The **no-causation rule** and the dataset-specific caveats stop a 14-row export being presented as proof. The final "one more piece of data" item gives you a next step.

## When not to use it

Only paste small datasets; beyond a few dozen rows, use a tool that runs code (a spreadsheet or a code-running AI) so the arithmetic is computed, not written out. Don't paste customer personal data into a chat tool without permission. For real significance testing, use a stats tool.

---

Canonical HTML: https://promptabide.com/bides/analyse-a-small-csv-three-findings-with-caveats
Agent guide: https://promptabide.com/llms.txt · https://promptabide.com/agent-instructions.md
Sitemap: https://promptabide.com/sitemap.xml