Data & Analysts
For business analysts and anyone who turns data into decisions — profile a dataset, write and explain a query, turn a CSV into a chart, and draft the stakeholder summary, all from the Claude Desktop chat with no SQL memorized.
Full worked systems, not one-off prompts — each walks the whole workflow end to end. Pick one and follow it.
Interview a dataset before you touch it
Open a file you've never seen and get the full picture — row count, what every column means, the date range, blanks, and duplicates — before you ask a single analysis question.
Build the metric dictionary every query and report inherits
Turn the metrics your team argues about — active, churn, revenue — into one dictionary: each with a plain-English meaning, the exact logic, an owner, and its edge cases, so every query, chart, and report finally counts the same thing.
Turn a vague request into an answerable analysis brief
Convert "can you look into churn?" into a one-page brief — the decision it informs, the precise questions, the exact metrics and sources, and what's out of scope — so you build the right analysis once instead of reworking it three times.
Write the query and learn the SQL
Get the exact query for the question you actually have, a line-by-line plain-English walkthrough of every clause, and a reconciliation against a known total — the answer and the understanding to defend it.
Turn two messy files into one trustworthy dataset
Standardize the keys, join on the right id, flag every row that doesn't match, and reconcile the result against the inputs — so the combined file is something you can actually build a report on.
Audit a dataset's quality before you trust it
Run the systematic checks — completeness, validity, uniqueness, referential integrity, freshness, and drift — and get a fit-to-use scorecard with severities, so you decide whether the data can feed a report or a decision on evidence, not on "it looks fine."
Find out why the number moved
Answer "why did it drop?" for real — compare the suspect period to a baseline, break it down by the right dimension, and rule out a broken query or dirty data before you blame the business.
Read out an A/B test honestly
Go from raw experiment data to a defensible ship decision — validate the test was sound, read the primary metric with its uncertainty, check the guardrails didn't regress, and resist the green number — so you ship on evidence, not on noise.
From a raw CSV to a chart you can paste into a slide
Pick the right view, generate the chart as an image file with clear titles and labels, and sanity-check the numbers behind it — so what lands in the deck is right, not just pretty.
Build a recurring analysis report, end to end
Profile the data, write and explain the queries, reconcile every result, chart the key views, and wrap it in a plain-English narrative with caveats — a report you re-run the same way every month so comparisons compound.
Stand up the team's self-serve analytics system
Assemble the metric dictionary, a library of saved and reconciled queries, the data contracts, a CLAUDE.md that teaches Claude your conventions, and scheduled monitors into one self-serve system — so a non-analyst can get a trustworthy answer and the analyst's time goes to real analysis.
Data & Analysts with Claude, in depth
The playbooks above are single jobs. These are the long reads behind them — how the work fits together across a quarter, what good looks like, and the judgement calls the prompts assume you have already made. Read in any order.
- Data Foundation: the dataset, the dictionary & the brief every analysis inherits Module 1 of the Data & Analysts deep dives — go past the recipe to master the three things every query, chart, and readout inherits: a profiled dataset you can trust, a metric dictionary that ends the 'two analysts, two numbers' problem, and an analysis brief that turns a vague ask into an answerable question. Build them for your own data. Free to sample.
- Experiments & charts: the readout and the picture that survive a second look Module 4 of the Data & Analysts deep dives — turn the experiment-readout and csv-to-chart playbooks into assessed judgment: read an A/B test honestly under pressure to ship, and turn a raw CSV into a chart that's verified before it ever reaches a slide. Build both for a real decision and check it against a rubric.
- Insight-in-a-Box — the Data & Analysts capstone The Data & Analysts capstone — integrate all five modules into one coherent analytics function for a single brand, proven end to end by taking one real investigation from raw export to a recurring, reconciled report. Checked against a master rubric that covers the artefacts and the judgement behind them.
- Quality & root cause: trust the number before you explain it Module 3 of the Data & Analysts deep dives — turn the foundation you built in M1 into two disciplines that protect every number leaving the folder: a fit-to-use quality audit before a dataset feeds something that matters, and a root-cause investigation that rules out a data artifact before it blames the business. Build it for your own data and check it against a real rubric.
- The Data Foundation toolkit: the files your team queries from The take-home pack for Module 1 — a ready data CLAUDE.md, fill-in templates for dataset-profile.md, metrics.md, and analysis-brief.md, plus the saved prompts that run every recurring task. Turn the foundation you wrote into shared infrastructure every query inherits.
- The query & join engine: turn a stakeholder ask into an auditable query, and two files into one trustworthy dataset Module 2 of the Data & Analysts deep dives — turn the dataset profile and metric dictionary you built in Module 1 into two production moves: a query a stranger could audit, and two messy files joined into one dataset you can actually report from. Build it for your own data and check it against a real rubric.
- The reporting system: the monthly report that compounds, and self-serve analytics Module 5 of the Data & Analysts deep dives — turn the recurring monthly report and the team's self-serve analytics into one orchestrated system: a report that comes out the same way every month and feeds straight into Finance's board pack, and a standing system — the dictionary, the query library, the data contracts, a CLAUDE.md, and scheduled monitors — that lets a non-analyst get a trustworthy answer without pulling you into every request.
The non-negotiables before any of this ships. Read the operating guide for the full data, brand, legal, and rollout playbook.
- Profile every new file and reconcile every result against a known total — a row count, a prior report, a finance figure. These two gates, at the start and the end, are what stop a confident wrong number from reaching a stakeholder.
- Define metrics once, in one place. Keep a single metric dictionary (
metrics.md) every query, chart, and report cites — and never let an analyst invent a definition in a chat. One source of truth for what 'active' and 'churn' mean is what keeps two analysts from being 'both right' with different numbers. - Have Claude write the query AND explain each line, and show its formula or steps for every total — so the number is reproducible and defensible by someone who wasn't in the conversation, not a black box.
- Audit data quality before it feeds anything high-stakes. A board figure, a recurring report, or a model deserves a real check — completeness, duplicates, referential integrity, freshness — because wrong numbers start at the input, not the output.
- Separate who prepares a number from who approves the conclusion it feeds. At enterprise scale a finding gets forwarded into decisions fast — keep the analyst who computed it distinct from the leader who acts on it, and attach the one confirming check to anything you report as a cause.
- Keep regulated or identifying data in a company-approved workspace, aggregate before you share or chart (revenue by region, not by named account), and approve each read in the 'Ask permissions' prompt — let Claude work the local copy, never pasted PII.
- Author bilingual readouts in their own language over one shared set of numbers. Write the Arabic narrative in Arabic for the right register — not machine-translated — but keep both versions citing a single reconciled appendix, so the English and Arabic packs can never quietly disagree on a figure.
- The deliverable is the stakeholder readout, not the query. Finish every analysis with the scoped, plain-English write-up for the audience and length asked for — a VP summary and an analyst's working notes are not the same document.
Questions people ask
- Do I need to know `SQL` or how to code to work as a business analyst with this?
- No. You drag a file into the Claude Desktop chat and ask in plain language; Claude writes the query and, if you ask, explains each line as it goes — so you learn the `SQL` while you get the answer. Nothing is typed into a terminal.
- How is this different from a free 'AI literacy' course for analysts?
- Those teach you what AI is and how to prompt a chatbot. This is hands-on with Claude Code Desktop — you point it at your real exports, requirements docs, and dashboards and actually do the analysis, reporting, and stakeholder write-ups an analyst is paid for.
- Can I trust a result without checking it?
- No — always reconcile a result against a known total, like a row count or a prior report, before you share it. A confident query can quietly drop rows; the Desktop visual diff and Claude's shown working let you catch that before a number reaches a stakeholder.
- Is it safe to put our data into Claude?
- Keep regulated or sensitive datasets in a workspace your company controls and let Claude work on the local copy through the Desktop file pane, approving each read in the 'Ask permissions' prompt. Profiling and querying an approved export this way is fine; pasting raw PII into any tool is not.