Generic AI data analysis sounds confident.Defensible analysis is built on five rules.
Upload a CSV to a generic AI prompt and you’ll get a clean-sounding summary that conflates correlation with causation, smooths over data-quality problems, and quotes numbers from segments with five rows as if they were reliable. The model isn’t lying — it’s doing what generic prompts ask of it. The fix is at the prompt layer.
The AI Data Analysis Starter Kit is one master prompt + eight focused prompts + a 24-question starter bank across Sales / Marketing / Finance / Operations. Every prompt enforces a five-rule honesty layer that keeps the output defensible. Paste, upload, run. One-time $79.
Generic AI analysis is confident. Defensible analysis is honest.
Modern AI assistants can absolutely run real calculations on your CSV — Claude’s analysis tool, ChatGPT’s data analysis, Gemini’s data mode all execute code in a sandbox and produce numbers from the actual file. The capability is in the box. The trap is the default prompt behavior: write a clean narrative, smooth over messy data, report segment-level numbers without flagging sample sizes, imply causation when only correlation is in the data.
The defaults are wrong because the generic prompt doesn’t tell the analyst to behave like a careful one. Confidence levels, small-sample flags, data-quality preflights, and correlation-vs-causation discipline aren’t natural defaults — they’re operator instructions. The starter kit makes those instructions the foundation of every prompt, not the part you remember to add.
"Channel X has a 78% win rate" sounds great until you notice it had 9 deals last quarter. The honesty layer flags any segment with too few rows and treats those findings as hypotheses, not results.
Generic prompts report that discounted deals win more often. The honesty layer forces the analyst to name plausible confounders — discounts may track desperate deals, not cause wins. Different conclusion, same data.
A 40%-null column gets averaged anyway and the answer looks confident. The honesty layer surfaces missing-values, duplicates, and likely entry errors BEFORE any insight, so they show up as caveats on the findings they affect.
Clear about the lane. No inflated promises.
- A paste-and-go prompt system for any AI assistant that runs code on uploads (Claude analysis tool, ChatGPT data analysis, Gemini data mode).
- One master prompt + eight focused prompts + a 24-question starter bank across Sales / Marketing / Finance / Operations.
- A five-rule honesty layer pasted into every prompt so the output stays defensible.
- Editable templates — customize the section headers, the chart preferences, the question bank for your industry.
- Designed for non-technical business owners — plain language, no jargon, anchored to decisions you actually face.
- 12 months of updates as new model capabilities ship and the prompts evolve.
- A BI / dashboarding tool. Use Looker / Mode / Metabase / Hex for ongoing dashboards.
- A database. Use Postgres / BigQuery / Snowflake / Sheets for storage.
- A statistical inference engine. For controlled experiments and A/B testing, use a tool built for that (Statsig, Optimizely, GrowthBook).
- A subscription. One-time $79 with 12 months of prompt updates.
- A replacement for an analyst on hard problems. The kit handles the 80% of analyses a generalist needs done; complex modeling still benefits from a specialist.
- A guarantee against bad decisions. The honesty layer makes the analysis defensible; the decision is still yours.
The IP. Pasted into every prompt.
Eight prompts are the surface; the five rules below are the kit’s actual IP. Every prompt enforces them. They’re the difference between an AI summary that sounds confident and an analysis that survives being shown to a CFO or a board.
Run real calculations on the uploaded file; don't estimate or eyeball numbers. If the tool can't run code, the prompt forces it to say so and label figures as approximate — never silently guess.
HIGH / MEDIUM / LOW with a one-line reason (sample size, data quality, variance). Findings without a stated confidence are findings you can't defend in the room. The rule forces the call.
If a segment or group has too few rows to support a conclusion, it's labeled and treated as a hypothesis rather than a result. The 5-deal-source that looks like 80% win rate is one bad week away from 40%; the kit makes that explicit.
Never imply X causes Y from data alone. The prompt forces the analyst to name plausible confounders worth ruling out. The classic discount-vs-win-rate trap (where discounts track desperate deals, not cause wins) only surfaces when this rule is in play.
Missing values, duplicates, mixed units, and likely data-entry errors get reported before any insight, and noted as caveats on findings they affect. Skipping this step is how you build a strategy on a column with 40% nulls.
Every analysis ends with a VERIFY BEFORE ACTING list — the specific things to double-check before betting on any finding. Not a generic disclaimer; the exact checks for the specific analysis. The closing rule that makes the rest actionable.
Nine sections. One dataset. One pass.
Upload a CSV, paste the master prompt, hit run. The output is structured into nine fixed sections so every dataset gets the same forensic pass — and so the output is comparable across datasets later.
2–3 plain sentences: what each row represents, the columns, the date range, the row count.
Missing values, duplicates, mixed formats, likely entry errors. Verdict: clean enough or fix first?
The single most important thing in this data, in one sentence, with its confidence level.
Each as: FINDING + numbers behind it + Confidence (HIGH/MED/LOW + why) + caveat if any.
Outliers and oddities, where they are, and whether each looks like real signal or a data problem.
What's moving over time and how segments compare — with the SIZE of each difference, not just direction.
2–4 visualizations that fit this data and the questions above, with one design note each.
The 1–3 actions this data actually supports, ranked. Not a wishlist; what to do this week.
The closing checklist — what to double-check before betting on any of the findings above.
Single-job runs when you don’t need the whole pass.
Sometimes you just need the data-health verdict. Or the chart recommendations. Or a conversational Q&A session. The eight focused prompts run a single job each — same honesty rules, less output to read.
Quality-only pass: row/column counts, missing values, duplicates, mixed-type columns, likely entry errors. Verdict: READY / FIX FIRST, with the exact fixes ranked.
Plain-English lay of the land. What each row is, column types, date range, min/max/median/mean per numeric column (flags mean-vs-median divergence as a skew signal), top values per categorical column.
3–5 most decision-relevant findings for a business owner. Each with numbers, confidence level, caveats. Ends with the single insight worth acting on first.
Outliers ranked by impact-if-real. Each labeled REAL SIGNAL or DATA ERROR with reasoning. Doesn't flag normal variation as anomaly.
Time-series + segment comparison, with the SIZE of each change (% and absolute). Small segments flagged. Seasonality called out where visible.
2–4 chart recommendations for a specific question. Each with type, column-to-axis mapping, design note. Flags any chart that would mislead given the data.
Conversational analyst. Plain-English questions, real-number answers with confidence + caveat. If a question can't be answered, says exactly what's missing instead of guessing.
Leadership-audience one-pager: headline + 3 key findings with confidence + 2–3 telling charts + ranked next steps + “what we're not yet sure about” note.
The hard part isn’t running numbers. It’s knowing what to ask.
24 starter questions across four functions, each tied to a metric and a watch-out (small sample, last-click attribution, mean vs median, classic correlation trap). Drop any one into the “Ask Your Data” prompt or use it to focus the master prompt.
"Which stage do deals die in most often?" · "Are discounts actually buying us wins, or just tracking desperate deals?" · "Who or what is concentration risk in our top 5 customers?"
"Which channel drives revenue, not just traffic?" · "Where do people actually drop off in the funnel?" · "Which content converted vs which just got clicks?"
"What's actually driving the change in profit — revenue or costs?" · "Where is margin leaking by product line?" · "Which customers or products are unprofitable when fully loaded?"
"Where's the actual bottleneck — and is it the longest stage or the queue?" · "What drives our support volume — what root causes create most tickets?" · "Where do errors or returns cluster?"
The integrity moat.
Exactly what you get for $79, and what you don’t.
- Master prompt (nine-section dataset profile).
- Eight focused prompts for single-job runs.
- Five-rule honesty layer pasted into every prompt.
- 24-question starter bank across Sales / Marketing / Finance / Operations.
- Quickstart guide for CSV prep and first run.
- Tuning notes covering long datasets, chunking, and "safe-to-edit vs load-bearing" sections.
- 12 months of prompt updates.
- A BI / dashboarding tool. Use Looker / Mode / Metabase / Hex.
- A database or warehouse. Use Postgres / BigQuery / Snowflake / Sheets.
- An experiment platform. Use Statsig / Optimizely / GrowthBook for A/B testing.
- A statistical inference engine for causal modeling. The kit handles correlation discipline; causal inference at scale needs a specialist tool.
- API access or automation. The prompts are paste-and-go.
- A subscription. One-time $79 with 12 months of updates.
Pairs naturally with the Meeting Intelligence System ($49) — meeting outputs become the input for a quick “what did the data say?” pass on whatever metric the meeting decided to act on.
For solo operators running a board / investor cadence, pair with the Solo Founder Skills Pack ($79) — the executive-summary prompt produces the data narrative that the board-update skill turns into an investor-grade memo.
For founders raising capital, pair with the Investor-Ready Metrics Pack ($89) — this kit handles the dataset-level honesty; the Metrics Pack handles the diligence-grade metric calculations and stage-aware benchmarks that go on top.
The questions non-technical operators actually ask before running their first analysis.
One master prompt that profiles a dataset end-to-end across nine sections (data health, headline, key insights, anomalies, trends & comparisons, recommended charts, next steps, verify-before-acting). Eight focused prompts for single jobs (Data Health Check, Exploratory Overview, Insight Extraction, Anomaly Detection, Trend & Comparison, Chart Recommendations, “Ask Your Data” Q&A, One-Page Executive Summary). A 24-question starter bank across Sales / Marketing / Finance / Operations to anchor your first analyses to real decisions. And a five-rule honesty layer pasted into every prompt so the output stays defensible.
Any assistant with a code/analysis capability that can actually compute on an uploaded file — Claude's analysis tool (the in-chat code execution that handles CSVs and spreadsheets directly), ChatGPT data analysis (formerly Code Interpreter), Gemini's data analysis mode, or your own local model with a Python runtime. The prompts are model-agnostic; the only requirement is that the tool can run real calculations on the file rather than estimating from a description. The honesty rules explicitly tell the model to say so if it can't run code rather than guessing — that's part of the design.
Yes. The kit is built for a non-technical business owner — the master prompt explicitly tells the analyst to write in plain language for a non-technical audience. The quickstart walks through prepping your file (cleaning column headers, fixing date formats, exporting from your CRM / ad platform / accounting tool to CSV), and the question banks give you the right questions to ask for your function. The hardest part of data analysis for a non-analyst isn't running numbers; it's knowing what to ask. The question banks solve that.
Without the honesty layer, AI data analysis produces confident-sounding findings that conflate correlation with causation, smooth over data-quality problems, and report numbers from small segments as if they were reliable. The five rules in the honesty layer (compute on actual data; state confidence on every finding; flag small samples; separate correlation from causation; surface data-quality problems first) get pasted into every prompt and produce output you can actually defend. The kit's IP is the discipline, not the prompts themselves — the prompts just encode it.
Anything that fits in a spreadsheet or CSV — sales pipelines, ad reports, P&L data, subscription data, support tickets, web/app analytics exports, inventory data, time-tracking exports, survey responses. The question banks cover Sales (deals, pipeline, CRM), Marketing (ads, web, email), Finance (P&L, subs, invoices), and Operations (tickets, fulfillment, inventory). For very large datasets that don't fit in one upload, the tuning notes cover the chunking pattern. For non-tabular data (PDFs, free-form notes), use a different tool first to structure it.
Yes — the prompts ship as editable templates. Edit any of the section headers, add a custom field, swap a chart recommendation, replace the question bank with one calibrated to your industry. The only sections to leave intact are the five honesty rules — those are what make the output trustworthy. The tuning notes explicitly mark safe-to-edit vs load-bearing sections.
Yes — that's its purpose. Once you've uploaded a dataset and started a session, the Q&A prompt sets up the analyst as a conversational partner. You ask plain-English questions (“Is our discount strategy actually buying us wins?”), it runs the actual numbers, gives the answer with a confidence level and a caveat if the data is thin. If a question can't be answered from the data, it says exactly what's missing instead of guessing. The question banks are designed to be dropped into this prompt as starter prompts that map to real decisions.
30-day no-questions refund. Run the master prompt on three of your real datasets. If the output isn't more defensible than what you'd get from a generic AI prompt — and if the confidence flags / small-sample callouts / data-quality warnings don't surface at least one finding you would have miscalled otherwise — email and we refund. Refunds across the catalog are countable on one hand; for data analysis specifically, the honesty layer earns its keep on the first dataset that has a small-segment trap or a data-quality issue.
Stop trusting confident summaries.
Start running defensible analysis.
One master prompt + eight focused prompts + a five-rule honesty layer + 24 starter questions across the four functions every operator runs. Paste, upload, run. Works in any AI assistant with a code-execution capability.
Sold by RedHub AI LLC · Secured by Stripe · redhub.ai