Data Analysis prompts

Help me understand this dataset before I analyse it.

DATA:
"""
[PASTE A SAMPLE, INCLUDING HEADERS]
"""

WHAT IT IS: [WHERE IT CAME FROM, WHAT IT MEASURES]
ROW COUNT: [TOTAL]
PERIOD COVERED: [DATES]
WHAT I WANT TO FIND OUT: [YOUR QUESTION]

Produce:

1. WHAT EACH COLUMN IS - your reading of each field: what it measures, its type, and its likely unit. Flag any column whose meaning is unclear from the name; ambiguous fields are where analysis goes wrong.

2. THE QUESTIONS TO ASK ABOUT THE DATA before analysing:
   - How was it collected, and what does that exclude? Every dataset has a population it cannot see.
   - What does a row represent, exactly? One customer, one order, one order line, one event?
   - Is it complete for the whole period, or does coverage change partway through?
   - Are there records that should be excluded: tests, internal users, cancelled, refunded, duplicates?

3. THE CHECKS TO RUN FIRST - specific to this data: row count, distinct counts on key fields, date range and gaps, null counts per column, and the distribution of the main numeric fields. Give the commands or formulas.

4. WHAT I EXPECT TO BE WRONG - based on the sample: likely duplicates, mixed formats, outliers, impossible values, and fields that are null more often than they should be.

5. THE DEFINITION PROBLEMS - where the analysis will depend on a definition that is not in the data. 'Active user' and 'churned' do not exist as facts; they are decisions. Name the ones my question requires.

6. WHETHER THIS DATA CAN ANSWER MY QUESTION - honestly. Frequently the answer is that it can answer a narrower question, or that a confounder makes the intended question unanswerable. Say so now rather than after the analysis.

7. WHAT IS MISSING that would make this analysis solid.

Do not analyse yet. Understand the data first.

Analyse this data.

DATA:
"""
[PASTE]
"""

WHAT IT IS: [DESCRIPTION AND SOURCE]
MY QUESTION: [WHAT YOU WANT TO KNOW]
WHAT DECISION THIS INFORMS: [WHAT YOU WILL DO WITH IT]
WHAT I EXPECT TO FIND: [YOUR PRIOR - be honest]

Produce:

1. THE DESCRIPTIVE PICTURE first - counts, ranges, distributions, and time coverage. Do this before any comparison. Most questions are answered here.

2. THE ANSWER TO MY QUESTION - stated directly, with the numbers that support it and the calculation shown.

3. THE CONFIDENCE - how strongly the data supports this. Consider: sample size, whether the difference exceeds normal variation, whether the period is representative, and whether the comparison is like for like. Say plainly if the answer is 'this data cannot tell you'.

4. WHAT WOULD EXPLAIN THIS OTHER THAN MY HYPOTHESIS - alternative explanations for the pattern. At minimum consider: a change in what was measured or how, a composition change (the mix of things being averaged shifted), seasonality, an outlier driving the aggregate, and survivorship (you only see what remained).

5. THE COMPOSITION CHECK - run this specifically. An overall average can move while every subgroup stays flat, purely because the mix changed. This is Simpson's paradox and it is common enough to check every time.

6. THE OUTLIER CHECK - is the finding driven by a few extreme records? Show the result with and without them.

7. WHAT I EXPECTED VS WHAT IS THERE - I told you my prior. Where the data contradicts it, say so directly. Where the data is consistent with it, note that consistency is weaker evidence than it feels.

8. WHAT THIS DOES NOT SHOW - the limitations, specific to this dataset.

9. THE DECISION - given what I said I would do with this, does the analysis support that action? If the finding is real but too small to matter, say so.

Rules:
- Show the arithmetic for anything I might need to check
- Never say 'significant' without meaning it statistically and showing the basis
- Correlation is not causation; say 'associated with' and name what would establish causation

Write analysis code.

LANGUAGE AND LIBRARIES: [Python/pandas / R / SQL / other]
DATA:
"""
[PASTE SAMPLE ROWS WITH HEADERS]
"""
FILE FORMAT AND SIZE: [CSV/Excel/database, and how large]
WHAT I WANT TO WORK OUT: [YOUR QUESTION]
OUTPUT I WANT: [numbers / table / chart / file]

Produce code that:

1. LOADS THE DATA with explicit types. Never rely on type inference for IDs, dates or anything with leading zeros. Specify dtypes and date parsing explicitly, and say why.

2. VALIDATES BEFORE ANALYSING - and prints the results so I can see them:
   - Row count loaded, versus expected
   - Null counts per column
   - Date range
   - Distinct counts on key fields
   - Any rows failing a basic plausibility check (negative quantities, dates in the future, impossible values)
   Stop and report rather than silently proceeding if something is wrong.

3. CLEANS EXPLICITLY - every cleaning step in its own line with a comment saying what and why, and printing how many rows it affected. Never drop rows silently; a filter that removes 40% of the data should be visible.

4. ANSWERS THE QUESTION - the analysis itself, with intermediate results printed rather than one chained expression. I need to be able to check the middle steps.

5. PRINTS THE ARITHMETIC - for any headline number, show the components. A single figure I cannot decompose is a figure I cannot defend.

6. HANDLES THE EDGE CASES - empty groups, division by zero, missing values in an average, and groups with too few rows to be meaningful.

Then, separately from the code:

7. THE ASSUMPTIONS the code makes about my data.
8. WHAT WOULD MAKE THE ANSWER WRONG - the data condition that would produce a plausible but incorrect result.
9. THE CHECK - a second, independent way to compute the headline number, so I can verify it.

Rules:
- Comment why, not what
- No chained one-liners for the main analysis; readable steps
- Do not use inplace operations or overwrite the source dataframe
- If the data is large enough that the approach matters, say so and adjust

Help me sanity-check this number.

THE CLAIM: [THE NUMBER AND WHAT IT SUPPOSEDLY SHOWS]
WHERE IT CAME FROM: [SOURCE]
HOW IT WAS CALCULATED: [IF STATED]
SUPPORTING DATA: [PASTE, if you have any]
CONTEXT: [WHAT DECISION DEPENDS ON THIS]

Produce:

1. THE DEFINITION CHECK - what exactly is being counted, over what period, for which population? Most wrong numbers are not calculation errors; they are definition mismatches. Name every ambiguity in how this is stated.

2. THE ORDER OF MAGNITUDE CHECK - is this number plausible at all? Work out roughly what it should be from independent knowledge and compare. A back-of-envelope estimate catches errors that detailed checking misses.

3. THE ARITHMETIC - if I gave you the calculation, redo it. If I gave supporting data, compute it yourself and compare.

4. THE COMMON ERRORS to check for in a number of this type:
   - Percentage of what? A different denominator changes everything.
   - Percentage change versus percentage point change - frequently confused, and the difference is large
   - Double counting, especially where records can appear twice
   - A rate quoted without its base, so 300% growth turns out to be from two to eight
   - Averages hiding a skewed distribution; the median may tell a different story
   - A period comparison where the periods are not equal length or are seasonally different
   - Currency, units or scale mismatches (thousands versus millions)
   - Survivorship: measuring only the ones that remained

5. WHAT WOULD MAKE THIS NUMBER TRUE, and what would make it false. Both.

6. THE QUESTIONS TO ASK the person who produced it, phrased precisely and without accusation. Usually: what exactly is in the numerator and denominator, what period, and what was excluded.

7. THE VERDICT - plausible / needs verification / likely wrong, with reasoning.

8. HOW MUCH IT MATTERS - given the decision, does the uncertainty change what you should do? Sometimes the number is unverifiable and the decision is the same either way.

Help me find where this change came from.

WHAT CHANGED: [THE METRIC, AND FROM WHAT TO WHAT]
PERIOD: [WHEN]
DATA:
"""
[PASTE - INCLUDE THE DIMENSIONS YOU CAN SEGMENT BY]
"""
WHAT I ALREADY SUSPECT: [YOUR HYPOTHESIS, if any]
WHAT CHANGED IN THE BUSINESS: [ANY KNOWN CHANGES]

Produce:

1. THE DECOMPOSITION - break the overall change down by each available dimension. For each segment: its value before, after, the change, and its contribution to the total change. Sort by contribution.
   Contribution, not percentage change, is what matters: a segment that doubled but is 1% of the total explains nothing.

2. THE CONCENTRATED OR DIFFUSE QUESTION - is the change driven by a few segments or spread across all of them? This distinction determines the whole investigation. Concentrated points to something specific; diffuse points to something systemic like a measurement change or a platform-wide effect.

3. MIX VERSUS RATE - separate two effects that look identical in an aggregate:
   - The rate within segments changed
   - The mix of segments changed while each segment's rate held
   Quantify each. This is the single most useful decomposition in business analysis, and it frequently reverses the apparent conclusion.

4. NEW VERSUS EXISTING - where applicable, split the change between new entities appearing, existing ones changing, and ones disappearing. An overall figure holding steady while half the base churned and was replaced is a very different situation.

5. THE TIMING - when exactly did it start? A sharp step points to a specific event or a measurement change; a gradual slope points to a trend. Say which this looks like.

6. AGAINST MY HYPOTHESIS - does the decomposition support what I suspected? If the change is spread evenly across all segments and I suspected one specific cause, that is evidence against me.

7. THE MEASUREMENT CHECK - before concluding anything about the business, rule out that the measurement changed: tracking, definitions, data collection, or a source system. A step change across every segment simultaneously is usually a measurement artefact.

8. WHERE TO LOOK NEXT - the segment or question that would most narrow this down.

Help me decide what to measure.

WHAT I AM TRYING TO ACHIEVE: [THE GOAL]
WHAT I CURRENTLY MEASURE: [EXISTING METRICS, or 'nothing']
WHAT DATA I CAN ACTUALLY COLLECT: [AVAILABLE SOURCES]
WHO WILL SEE THESE NUMBERS: [AUDIENCE]
WILL ANYONE BE INCENTIVISED ON THEM: [yes / no]

Produce:

1. THE OUTCOME METRIC - the one number that best represents the goal being achieved. Usually there is exactly one, it is hard to measure, and it moves slowly. Name it even if it is impractical, because everything else is a proxy for it and you need to know what you are proxying.

2. THE PROXY METRICS - what you can actually measure that correlates with the outcome. For each: how well it proxies, and where the correlation breaks.

3. THE LEADING INDICATORS - what moves before the outcome does, so you can steer rather than report.

4. THE GUARDRAILS - what must not get worse while you optimise the main metric. Every metric has a way of being improved that damages something else. Name the specific harm for each proposed metric.

5. THE GAMING ANALYSIS - if anyone is incentivised on these, what would they do? List three specific behaviours per metric that would improve the number without achieving the goal. If people will be incentivised on these, this is the most important section.

6. THE DEFINITION - for each metric you recommend: exactly what counts, what does not, over what period, and how it is calculated. Ambiguity here becomes an argument later.

7. WHAT NOT TO MEASURE - vanity metrics that will be reported and mean nothing. Say why each is a distraction.

8. THE MINIMUM SET - the fewest metrics that would tell you whether this is working. Usually three: one outcome or proxy, one leading indicator, one guardrail. More than five and nobody looks at any of them.

9. THE BASELINE - what to capture now, before you start, or you will not be able to tell whether anything changed.

Rules:
- Prefer rates and ratios over raw counts, which grow with the business regardless of performance
- Every metric needs a denominator that makes it comparable over time

Help me interpret this result.

THE RESULT:
"""
[PASTE - THE NUMBERS, P-VALUES, CONFIDENCE INTERVALS, SAMPLE SIZES, WHATEVER YOU HAVE]
"""

WHAT WAS BEING TESTED: [THE QUESTION]
HOW THE DATA WAS COLLECTED: [METHOD]
WHAT DECISION THIS INFORMS: [WHAT YOU WILL DO]
WHAT I THINK IT MEANS: [YOUR READING]

Produce:

1. WHAT IT LITERALLY SAYS - in plain words, precisely. If there is a p-value, state what it actually means, which is not what most people think: the probability of seeing a result at least this extreme if there were genuinely no effect. It is not the probability that the hypothesis is true.

2. THE EFFECT SIZE - more important than significance and usually omitted. How big is the difference, in units that matter to the decision? A statistically significant effect can be far too small to act on.

3. THE CONFIDENCE INTERVAL - if given, what range of true values is consistent with this data. If the interval spans from 'barely worth doing' to 'transformative', the honest answer is that you do not yet know.

4. THE SAMPLE SIZE QUESTION - was this large enough to detect the effect it claims? An underpowered study that finds an effect tends to overstate its size.

5. WHAT COULD PRODUCE THIS RESULT OTHER THAN A REAL EFFECT:
   - Multiple comparisons: how many things were tested? Test twenty and one will look significant by chance.
   - Selection: who was included, who was excluded, and who dropped out
   - Timing and seasonality
   - A confounder correlated with both the treatment and the outcome
   - Stopping when the result looked good

6. WHAT IT DOES NOT SHOW - especially: whether it generalises beyond this sample, whether the relationship is causal, and whether it will persist.

7. MY READING, ASSESSED - I said what I think it means. Where I am over-reading, say so directly.

8. THE DECISION - given what I said I would do, is this result strong enough to act on? Three honest options: act, gather more data, or act anyway because the cost of being wrong is low. Say which and why.

Do not make the result sound more conclusive than it is.