10 Claude Prompts for Data Analysis That Actually Work

You are a senior data analyst doing a first-pass profile of a new dataset.

Here is the dataset (or a sample of it):
[paste CSV sample / schema]

Profile it before any analysis:
1. What does each column appear to represent? Flag ambiguous ones.
2. Data quality issues: nulls, duplicates, impossible values, mixed types, suspicious outliers.
3. Which columns are dimensions vs. metrics?
4. What questions could this dataset credibly answer — and which questions would it NOT support?
5. The three checks you'd run before trusting any conclusion from this data.

Be specific: reference actual column names and example values.

I need to answer this business question: [question, e.g. "why did signups drop 18% in June?"]

Available data: [describe tables/columns or paste sample]

Work like an analyst, not a search engine:
1. List 4–6 plausible hypotheses that could explain this, ordered by prior likelihood.
2. For each hypothesis, state exactly what pattern in the data would confirm or kill it.
3. Specify the queries/cuts needed to test each one (write them if you have the schema).
4. State what result would make you say "we don't have the data to know."

Do not conclude anything yet — this is the analysis plan.

Write a [PostgreSQL / BigQuery / Snowflake] query for this request:

Request: [e.g. "monthly retention by signup cohort for the last 12 months"]

Schema:
[paste CREATE TABLE statements or column lists]

Requirements:
- Use CTEs with descriptive names, one logical step per CTE.
- Comment any non-obvious business logic.
- State every assumption you made about the schema (grain, timezone, soft-deletes, duplicates).
- List 2–3 sanity checks I should run to verify the output is right.
- If the request is ambiguous, ask me the clarifying question INSTEAD of guessing.

Here is an analysis/claim I'm about to present:
[paste your finding, e.g. "users who enable feature X retain 2.3x better"]

And the underlying numbers:
[paste the data or summary stats]

Attack it like a skeptical reviewer:
1. Is the sample size adequate? Show the rough math.
2. Correlation vs. causation: what confounders could produce this pattern?
3. Survivorship, selection, or seasonality effects that could explain it?
4. Would this survive a segment split (by cohort, geography, plan tier)?
5. Verdict: "solid", "directionally useful", or "not supported" — and the one test that would upgrade it.

I need to visualize this for [audience, e.g. "the exec team"] to support this point: [the takeaway, e.g. "growth is concentrated in two regions"].

Data shape: [describe, e.g. "monthly revenue for 8 regions over 24 months"]

Recommend:
1. The single best chart type, and why it beats the two obvious alternatives.
2. What goes on each axis, what gets aggregated, what gets dropped.
3. The title written as the takeaway (a sentence, not a label).
4. Two mistakes people typically make with this chart type.

Then generate the chart code in [Python/matplotlib, plotly, or React/Recharts].

This metric moved and I need to know why: [metric, direction, magnitude, when — e.g. "checkout conversion fell from 3.1% to 2.4% starting May 14"].

Context: [product, recent changes, seasonality you know about]
Data I can pull: [tables/dimensions available]

Run a structured root-cause investigation:
1. Segment first: which cuts (device, geo, channel, browser, plan) would localize the drop? Order them by diagnostic power.
2. Timeline: what internal events (deploys, pricing, campaigns) and external events should I check against May 14?
3. Denominator check: did the metric fall because the numerator dropped or the denominator grew?
4. Give me the decision tree: "if segment X shows it, look at Y next."

Design a cohort analysis for: [question, e.g. "is our onboarding redesign improving week-4 retention?"]

Product context: [what the product is, what 'active' means, when the change shipped]

Specify precisely:
1. Cohort definition (what event, what grain — day/week/month).
2. The activity metric and its exact definition (avoid vanity definitions — justify your choice).
3. The comparison that isolates the change (pre/post cohorts, and what contaminates them).
4. How many periods until the answer is trustworthy.
5. A mock table of what the output should look like, with realistic fake numbers so I can sense-check the shape.

Act as my pair-programmer for exploratory data analysis in Python (pandas).

Dataset: [path or paste sample + column descriptions]
Goal: [what you want to learn]

Write a single, runnable analysis script that:
1. Loads and validates the data (assert expected columns, log rows dropped and why).
2. Runs the analysis in clearly separated, commented steps.
3. Prints intermediate results so I can verify each step, not just the ending.
4. Ends with a plain-English summary of findings printed to stdout.

Style: prefer readable over clever. No functions longer than a screen. If a step involves a judgment call (imputation, outlier handling), stop and flag it as a comment starting with # DECISION:.

Turn this analysis into an executive summary:

[paste your analysis, tables, or notebook conclusions]

Audience: [e.g. "CFO — numerate, skeptical, 90 seconds of attention"]
Decision at stake: [what will be decided based on this]

Produce:
1. A one-sentence headline stating the finding (not the topic).
2. Three supporting bullets, each with a number.
3. The caveat that matters (just one — the one that could change the decision).
4. The recommendation, phrased as an action with an owner and a timeframe.

Cut everything that doesn't serve the decision. No methodology unless it changes the conclusion.

Our team disagrees about how to define this metric: [metric, e.g. "active user"]

Positions:
[summarize each camp's definition and reasoning]

Business context: [what the metric drives — goals, comp, board reporting]

Arbitrate:
1. What behavior does each definition actually incentivize? Where would each be gamed?
2. What does each definition hide? Give a concrete scenario where they'd tell opposite stories.
3. Industry norms for this metric in [industry], and where norms shouldn't apply to us.
4. Recommend one definition, one guardrail metric to watch alongside it, and the migration note for historical comparisons.