Three master prompts for a full-day analytics workshop. Each is dataset-agnostic: paste it above any dataset and the model reasons like a seasoned analyst, then turns around and attacks its own conclusion. Tap copy on any card, paste straight into Claude, ChatGPT, Gemini, or Copilot.
Workshop lecturerSteve Sands
The three prompts
What is going on
This is the day's centrepiece, and it is deliberately dataset-agnostic: it works on the mystery set, the Detroit taxis, or anything else with no editing. It makes the model reason the way a twenty-year analyst does, then turn around and try to demolish its own best finding.
The marvel in an analytics room is never the AI crunching numbers; everyone expects that. It is watching the machine commit predictions before it has seen the data (the cold read), and then attack its own headline conclusion by name (the red team). That second move tends to make a class go quiet.
The teaching point to frame the day around
A novice points AI at data and asks for an answer; the output looks like analysis and is often wrong in ways they cannot see. The lesson is not "AI does the analysis". It is "AI can hold the analyst's discipline, and your job is to supply the discipline it is missing".
How to run it live, three beats do the work:
The cold read. Run step 1 only, on a dataset the model has not seen. Have it commit its three predictions on screen, then run step 2 and reveal which held. Watching a prediction land feels uncanny.
Stop the room at step 6. This is the gasp. Few have seen an analyst voluntarily attack their own headline and name the exact bias that could be fooling them. Pause and ask: "When did a colleague last do this, unprompted?"
The ninety-second narrative. The final compression from raw rows to a decision a leader can act on is the skill the whole day is trying to build. Show it in one pass.
Dials you can turn
Sharpen the skeptic (step 6): "Assume a hostile reviewer will publish a takedown. Write their strongest paragraph, then respond to it." · Force the trade-off (step 7): "Now give the opposite recommendation and the conditions under which it would be correct." · Effect size over significance (step 4): "Treat any p-value as suspect until you have told me how large the effect is and whether a practitioner would care." · Slower group: run one step at a time and have tables predict what the model will say before revealing it.
The prompt
You are a senior data analyst with twenty years of experience, the kind who has learned to distrust the first answer. I am going to give you a dataset. Work through it in the order below, and narrate your reasoning out loud at every step so I can watch how an experienced analyst actually thinks. Do not skip ahead.
1. COLD READ. Before you look at a single summary statistic, tell me what this data appears to be, who likely produced it and why, and the three patterns you would bet money on finding given the domain. Commit to those predictions in writing.
2. INTERROGATE. Now examine the data. What is its shape, what is missing, what looks too clean, what looks suspicious? Name anything that would make you cautious: gaps, outliers, suspected data-entry artefacts, definitions that could mean two things. Tell me where your cold-read predictions held and where they broke.
3. QUESTIONS WORTH ASKING. Generate the ten questions this dataset can answer. Then rank them not by how easy they are to answer, but by how much a real decision would change if you knew the answer. Keep the top three and tell me why those three.
4. ANSWER WITH CONSEQUENCE. Answer your top three. For each, give the direction, the size of the effect in plain units a non-expert would understand, and who it matters for. No statistic without a "so what".
5. THE THING A JUNIOR WOULD MISS. Surface the one non-obvious insight that only appears when you connect two parts of the data that do not sit next to each other.
6. RED TEAM YOURSELF. Take your single strongest conclusion and try to destroy it. What would have to be true for it to be wrong? Walk through the classic traps by name: confounding, selection and survivorship effects, reversed causation, Simpson's paradox, a sample that does not represent the population. Tell me how confident you remain after the attack, and why.
7. THE DECISION. State the one recommendation you would put in front of a decision-maker, your confidence as a percentage, and the single piece of additional data that would most change your mind.
Finally, rewrite the whole thing as one short narrative a busy executive would read in ninety seconds, with no jargon.
What is going on
This one turns the model into an elite, multi-disciplinary data team: a Chief Data Scientist, a Senior BI Analyst, and a Master Data Storyteller working as one. It then forces that team through four sealed phases and forbids it from moving on until each is complete in its reasoning.
Phase 1, Structural Audit & Hidden Signals. Finds the three non-obvious patterns a surface-level pivot table would miss, and states its assumptions.
Phase 2, the "So What?" Layer. Translates each signal into a business implication and the cost of ignoring it.
Phase 3, Counter-Intuitive Friction. The devil's advocate. It challenges its own findings, names a bias or confounder, and says what extra data would settle it.
Phase 4, Executive Deliverables. A headline, a clean Markdown dashboard table, and a roadmap of three recommendations ranked by effort vs reward.
Why it lands
It avoids generic "here is a summary of your columns" output, the Phase 3 devil's advocate models the exact skepticism a real data scientist must hold, and the result arrives presentation-ready rather than as a wall of text. Students paste it alongside a CSV or Excel upload, or a clean data sample.
The prompt
You are an elite, multi-disciplinary data analytics team consisting of a Chief Data Scientist, a Senior Business Intelligence Analyst, and a Master Data Storyteller.
Your objective is to analyze the attached dataset and deliver a high-impact, executive-level data briefing.
Please process the data through the following four distinct phases. Do not move to the next phase until the previous one is completed in your reasoning.
### Phase 1: Structural Audit & Hidden Signals
- Conduct a rigorous audit of the data structure.
- Identify the 3 most compelling, non-obvious patterns, correlations, or anomalies that a standard surface-level pivot table would completely miss.
- State any baseline assumptions you are making about the data context.
### Phase 2: The "So What?" Layer (The Business Intelligence Lens)
- Translate those 3 hidden signals into critical business or operational implications.
- For each signal, answer: "Why does this matter to a stakeholder, and what is the cost of ignoring it?"
### Phase 3: Counter-Intuitive Friction (The Devil's Advocate)
- Challenge your own findings. Identify one potential data bias, confounding variable, or alternative explanation that could prove our Phase 1 conclusions wrong.
- Suggest what *additional* data would be needed to validate or disprove this friction point.
### Phase 4: Executive Deliverables
Present your final output in a highly professional, scannable format using the following structure:
1. **The Headline:** A single, high-impact sentence summarizing the ultimate breakthrough from the data.
2. **The Executive Dashboard (Text-Based):** A clean Markdown table featuring the 3 core metrics/KPIs that matter most, their current state, and their trajectory.
3. **Strategic Roadmap:** 3 concrete, data-backed recommendations ranked by operational impact (High/Medium/Low effort vs. High reward).
What is going on
This is the full-day spine in a single prompt. Its job is not just to analyse the data, but to teach the student how to think like an analyst. It refuses to invent data, asks for anything missing, and states limitations clearly.
It walks seven stages, the arc of the whole workshop: understand → question → analyse → interpret → report → critique → teach. It ends by producing a professional report for a senior, non-technical manager, then turning around to explain what the student should have learned from the exercise.
The rule to give students alongside it
"You are not allowed to accept the AI's answer until you have challenged one assumption, checked one caveat, and improved one question." That single rule turns the exercise from "AI writes my report" into "AI makes me a better analyst".
The prompt
You are my AI analytics partner for a full-day data analytics workshop.
Your job is not just to analyse the data, but to teach me how to think like an analyst.
I will provide a dataset, a data dictionary, or a sample of rows. Do not invent data. If something is missing, ask for it. If the dataset has limitations, say so clearly.
Work through the analysis in seven stages.
STAGE 1: Understand the Dataset
First, inspect the dataset and explain in plain English:
- What each column appears to represent
- What the unit of analysis is, for example customer, trip, transaction, employee, product, order, visit, claim
- What questions this dataset is naturally able to answer
- What questions it cannot answer safely
- Any obvious risks, missing values, outliers, duplicated rows, strange categories, or data quality issues
STAGE 2: Generate Better Questions
Create 12 strong analytical questions from the dataset:
- 4 descriptive questions: what happened?
- 4 diagnostic questions: why might it have happened?
- 2 predictive questions: what might happen next?
- 2 decision questions: what should someone do about it?
Rank the questions by business value and feasibility.
STAGE 3: Build an Analysis Plan
Choose the 5 best questions and design an analysis plan for each one.
For each question, give:
- The business reason it matters
- The columns needed
- The method to use
- The chart or table that would best reveal the answer
- The likely caveats
- What a weak conclusion would sound like
- What a strong evidence-based conclusion would sound like
STAGE 4: Do the Analysis
Now analyse the data.
For every finding:
- Give the result
- Explain how you got it
- State whether it is an observation, an inference, or a recommendation
- Mention any uncertainty or limitation
- Suggest one follow-up analysis that would strengthen the conclusion
Do not overclaim. Do not confuse correlation with causation.
STAGE 5: Find the Story
Turn the analysis into a story.
Give me:
- The headline insight
- Three supporting findings
- One surprising or counterintuitive result
- One thing that looks important but may be misleading
- One decision the organisation could make from this
- One thing the organisation should not decide yet because the evidence is insufficient
STAGE 6: Create the Report
Write a professional analytics report with this structure:
1. Executive Summary
2. Dataset Overview
3. Key Questions
4. Main Findings
5. Visuals to Include
6. Business Interpretation
7. Risks and Limitations
8. Recommendations
9. Further Analysis
Write it for a senior manager who is intelligent but not technical.
STAGE 7: Teach Me
Finally, explain what I should learn from this exercise as a developing analyst.
Include:
- The three most important analytical habits this dataset teaches
- The biggest mistake a beginner might make with this data
- How AI helped
- Where human judgement was still necessary
- Three ways to improve the analysis if we had more time