Computing a number is free; knowing whether it means what you think it means is not. Twenty prompts across eight functions that work from your data and never invent a figure.
The hard part was never the calculation
Computing a number is free. It has been free for years. What is not free is knowing whether the number means what you think it means — and that gap is where every prompt in this pack lives.
An arithmetically perfect figure can still be wrong in at least five ways before anyone acts on it. The metric was defined loosely, so two teams compute it differently. The sample excluded the people who churned. The comparison period contained a holiday. The lift is inside the noise. Or the number is correct and the decision it triggers does not follow from it.
None of those are computation errors. Every one of them is caught by process, not by a better query — which is why this is a pack of methods rather than a pack of formulas.
This is a lens, not a department
The twenty prompts tagged here come from eight different parts of the site: product, marketing, PR, social, business finance, data analysis, customer success and nonprofits. That is deliberate. Measurement is not something an analytics team does on behalf of everyone else; it is a discipline each function needs in its own vocabulary.
What they share is a refusal that runs through all of them: they do not invent numbers. Every prompt here works from data you supply. Where an input is missing, the output says so rather than filling the gap with a plausible-looking figure. That constraint is the whole reason these are usable on real work — see the research pack for why a confident, well-formatted number deserves more suspicion than a hedged one.
Decide what to measure before you can measure it
Almost every unusable dashboard traces back to a definition nobody wrote down.
Choose the Few Numbers You Check Every Week is the right starting point for a business, because the failure mode is not too few metrics — it is forty, none of which anyone acts on. A number nobody would change a decision over is not a metric, it is decoration.
Define Success Metrics and Guardrails for a Feature adds the half that gets skipped: the guardrail. Any metric can be moved by damaging something else, and a success criterion without a stated thing-that-must-not-get-worse is an invitation to do exactly that. From product management.
Design an Event Tracking Plan Before You Build the Feature is the highest-leverage prompt on this page and the one most often reached for too late. Instrumentation designed after launch cannot answer questions about the launch. You get one chance to collect the data, and it is before the code ships.
Write an Analysis Plan Before You Pull the Data commits to the method and the decision rule in advance. It exists because of a specific failure: an analysis that begins without a stated rule quietly becomes a search for the cut of the data that supports what you already wanted. Naming the rule first is the only reliable defence, and it costs ten minutes.
Get the data into a state you can trust
Clean and Validate a Messy Dataset and Reconcile Two Data Sources That Disagree handle the part of analytics that consumes most of the time and appears in none of the slides.
The reconciliation prompt earns its place because of what it refuses to do: two systems reporting different revenue figures usually disagree for a reason — different timezone boundaries, different treatment of refunds, different definitions of active
— and finding the reason is the actual work. Splitting the difference is not reconciliation, it is averaging two things you do not understand.
Work out what happened
Work Out Why a Number Moved is the request every analyst receives most often and the one most likely to produce a confident wrong answer. It decomposes the movement — mix, volume, rate, seasonality, a change you shipped — before reaching for a story, because the first plausible explanation is very rarely the whole one.
Check Whether a Difference Is Real or Just Noise is the prompt that should run before any of those explanations are offered. A great deal of organisational energy goes into explaining variation that is not there.
Design an Experiment or A/B Test forces the sample size and the stopping rule up front. Stopping a test when it first looks significant is the most common way to generate a result that does not replicate, and it feels like diligence while you do it.
Run a Cohort and Retention Analysis and Segment Customers From Behavior Not Intuition both replace an aggregate with a structure. An overall retention rate can be flat while every cohort inside it is deteriorating; a persona invented in a workshop can survive years without anyone checking it against what people do. From data analysis.
Make it usable by someone else
Sanity-Check an Analysis Before You Share It reads your finished work the way a sceptical reader would — definitions, sample, confounds, and whether the conclusion actually follows — and separates what is established from what is inferred. Run it as the last step before anything reaches a decision-maker.
Design a Dashboard People Will Actually Use starts from the decisions the dashboard should support. Most dashboards are built from what is easy to query and are consulted enthusiastically for a fortnight.
Measurement inside each function
Marketing measurement is mostly a fight against attractive numbers that mean nothing. Report Social Performance Without Vanity Metrics and Measure PR Without Counting Clippings both replace volume with something connected to an outcome — see social media and PR.
Find Out Which Products or Services Actually Make Money applies the same treatment to revenue, which is the most misleading number in most businesses because it is the one everyone already looks at. From business finance.
Measure Whether Customers Are Actually Using What They Bought is the leading indicator behind most churn, and it is visible months before the renewal conversation — customer success.
Design an Affiliate Program That Pays for Incremental Sales and Police an Affiliate Program and Cut the Bad Actors turn on the hardest question in attribution: whether the sale would have happened anyway. Paying commission on demand you already had is the default state of most affiliate programmes. See partnerships.
For nonprofits the pressure runs the other way — toward claiming more than the evidence supports. Build a Theory of Change That Survives a Funder's Questions, Write the Grant Report Your Funder Will Actually Read and Report Your Impact Without Overclaiming hold the line between what you did, what changed, and what you can actually attribute to yourself.
Where this stops
These prompts organise measurement; they do not perform it. They have no access to your warehouse, your analytics platform, or your numbers, and they will not produce a statistic, a benchmark, or an industry average — anything they generated there would be invention dressed as data.
They also cannot tell you that your instrumentation is lying. If an event fires twice, a filter silently drops a segment, or a join duplicates rows, every method here will run cleanly on wrong inputs and produce a well-structured wrong answer. Validate the pipeline; no framework substitutes for that.
And where the stakes are regulatory, clinical, financial-reporting or safety-related, statistical work needs a qualified professional and a review process, not a prompt.
Sources
- Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala and Edwin Zhang, Why Language Models Hallucinate, September 2025 — training and evaluation reward confident guessing over expressed uncertainty, which is why every prompt in this pack is built to work from supplied data and to flag what is missing rather than fill it in