your codeyours modelno accuracy Benefited, knowing that I might really make me difference In your group but then you struggle above share What is the findings with your team and your participants?
That is the most common feeling between data scientists and ml engineers.
In this article, I share my travel visits, work flow, and small curriculum tactics, sometimes abstract, model output sharp and clear Business accounts People actually care.
If you work with participants or Managers They do not live in writing all day, this is for you. And like some of my directors, I'll keep it activated and copied to cheat.
This article is a third and last of the 3 topics in relation to instant data science engineering.
This page End-until the end Data Science Prompt Engineering Series is:
👉 All motivation In this article are available at store of this article as a cheating paper 😉
In this article:
- Why llms is a change of storytelling data
- LifeCycle Communication, reinstated with llms
- It stimulates documents, devaps, and interactions of stakeholders
- The cheat sheet of Coupter
1) Why is llms a change of storytelling data for storytelling
Llms mix well hand reference same Reasoning. In fact, that means they can:
- Rephrase Tricky metrics in an obvious English (or other language),
- High Draft – Level condensation in seconds, and
- practice tone including format In any listener – the Board, Product, Officer, call it.
The original research shows that GPT style models actually can increase obedient A Members which is not technical Students in double digits. That is a great jump compared to just looking at charts or graphs.
And because llms “Talk to participants“They help you to prevent decisions without dripping people in Jargon.
If quick engineering sounds like hype earlier, here is a real edge: Clear NewsFew meetings, they bought immediately.
2) Communication LifeCycle, reinstated with llms
After training to check the model, you will probably:
- Translate the effects of model (shape, coefficients, confused matric).
- Summarize Edge and hit capeats.
- Executive Framework Executive Strient, slid documents, and “you can do next.”
- Make Memos and standard Demos.
- Close the loop with the rapid issues and quick updates.
Now: think a helper the first draft, explains OffsIt costs the missing context, and is always the word consistent across the authors.
That's what the llms can be, if you move them well!
3) It also motivates interpreting patterns, reporting, and participating participants
3.1 The most important narrative & feature
Good practice: Feed the model in orderly table and ask for a ready convenient summary for you and actions.
## System
You are a senior data storyteller experienced in risk analytics and executive communication.
## User
Here are SHAP values in the format (feature, impact): {shap_table}.
## Task
1. Rank the top-5 drivers of risk by absolute impact.
2. Write a ~120-word narrative explaining:
- What increases risk
- What reduces risk
3. End with two concrete mitigation actions.
## Constraints & Style
- Audience: Board-level, non-technical.
- Format: Return output as Markdown bullets.
- Clarity: Expand acronyms if present; flag and explain unclear feature names.
- Tone: Crisp, confident, and insight-driven.
## Examples
- If a feature is named `loan_amt`, narrate it as "Loan Amount (the size of the loan)".
- For mitigation, suggest actions such as "tighten lending criteria" or "increase monitoring of high-risk segments".
## Evaluation Hook
At the end, include a short self-check: "Confidence: X/10. Any unclear features flagged: [list]."
Why apply: → Emposexicomination → act. The participants receive “so?” Not just bars on the chart.
3.2 confusion-matrix clarification
Think your project is about Discovery Discovery with a finance platform.
He has trained a good model, your accuracy and scores look good, and you feel proud to do. But now it comes part when you need to explain those results groupor worst, in a full room participants They don't really understand with metric models.
Here is a valid table that describes confused words – Matrix's simple structures of English:
| Metric | Translation of plain-english | Prompt Snippet |
|---|---|---|
| Lie | “Appreciate but not actually fake” | Describe FP as review costs. |
| False Unfortunately | “You missed a real fake” | Fnline Fn as Loss of Income / Accident Exposure. |
| Well doing | “How many warning is good” | Mix with QA false alarms. |
| Remember | “How many real cases have caught” | Use analoom “fishing-net fishing. |
Immediately to describe the results in model easily
## System
You are a data storyteller skilled at explaining model performance in business terms.
## User
Here is a confusion matrix: [[TN:1,500, FP:40], [FN:25, TP:435]].
## Task
- Explain this matrix in ≤80 words.
- Stress the business cost of false positives (FP) vs false negatives (FN).
## Constraints & Style
- Audience: Call-center VP (non-technical, focused on cost & operations).
- Tone: Clear, concise, cost-oriented.
- Output: A short narrative paragraph.
## Examples
- "False positives waste agent time by reviewing customers who are actually fine."
- "False negatives risk missing real churners, costing potential revenue."
## Evaluation Hook
End with a confidence score out of 10 on how well the explanation balances clarity and business relevance.
3.3 ROC & AUC-Makes Trade Concrete
ROC scores and AUC Scores are one of your DSS Metrics, good by checking model performance, but usually far from business discussions.
To do real things, a tie model sensitivity including clarification In real business ends: as time, money, or responsibility.
Dream:
“Highlight the trade-off between 95% sensitivity and marketing cost; suggest a cut-off if we must review ≤60 leads/day.”This type of framing changes the invisible concrete metrics, working decisions.
3.4 regression metric cheat-sheet
When working with the Regression models, the metrics may feel as a group of random characters (MAE, RMSE, R²). It is good to understand the models, but it's not so good to tell stories.
That is why it helps reduce these numbers using simple business accountogies:
| Metric | A businessy business | One-liner template |
|---|---|---|
| Parent | “Randoms are closed at each scale” | “Our $ 2 mae says the normal average error is $ 2.” |
| RMSE SYTH | “The penalty grows with a great missing” | “RMS 3.4 → Arare now.” |
| ² | “Sharing Differences Explain” | “We picked up 84% of priced drivers.” |
💥 Don't forget to check Part 2 of this series, where you will learn to improve your own model including Engineering feature.
4) Summing EDA-and CAPEATS Top before
EDA is when the real job of real removal begins. But let's face: these automatically generated reports (such as pandas-profiling or summary of JSOs) can be difficult.
Soon the following helps to change the EDA's output into short and humanity summary.
Directed EDA clan (Pandas-Prodi or JSON summoned in, short):
## System
You are a data-analysis narrator with expertise in exploratory data profiling.
## User
Input file: pandas_profile.json.
## Task
1. Summarize key variable distributions in ≤150 words.
2. Flag variables with >25% missing data.
3. Recommend three transformations to improve quality or model readiness.
## Constraints & Style
- Audience: Product manager (non-technical but data-aware).
- Tone: Accessible, insight-driven, solution-oriented.
- Format:
- Short narrative summary
- Bullet list of flagged variables
- Bullet list of recommended transformations
## Examples
- Transformation examples: "Standardize categorical labels", "Log-transform skewed revenue variable", "Impute missing age with median".
## Evaluation Hook
End with a self-check: "Confidence: X/10. Any flagged variables requiring domain input: [list]."
5) Top Summary, visual recognition and slider accounts
After the data and generation of understanding, there is one last challenge: telling your data story On the way of decision makers in fact care about.
Snapshots frame
- Simple Summary Reference: Intro, key points, recommendations (words ≤500).
- Summary of news of the story: Great points, important statistics, Trend Lines (Names ≈200).
- Church “Prompt Prompt”: Two short sections + “the following steps” characters.
ComotionCite Prompt
## System
You are the Chief Analytics Communicator, expert at creating board-ready summaries.
## User
Input file: analysis_report.md.
## Task
Draft an executive summary (≤350 words) with the following structure:
1. Purpose (~40 words)
2. Key findings (Markdown bullets)
3. Revenue or risk impact estimate (quantified if possible)
4. Next actions with owners and dates
## Constraints & Style
- Audience: C-suite executives.
- Tone: Assertive, confident, impact-driven.
- Format: Structured sections with headings.
## Examples
- Key finding bullet: "Customer churn risk rose 8% in Q2, concentrated in enterprise accounts."
- Action item bullet: "By Sept 15: VP of Sales to roll out targeted retention campaigns."
## Evaluation Hook
At the end, output: "Confidence: X/10. Risks or assumptions that need executive input: [list]."
6) Tone, clarity, and formatting
Get information and conclusions. It is time to make them clear, confident, and easy to understand.
Information scientists know how you say something is more important than what you say!
| Tool / Quick | What is | General use |
|---|---|---|
| “The writer of Tin Rewriter” | Formal ↔ very good, or “ready” | Customer updates, EXT Memos |
| Hemandayy-style Edit | Reduce, Punch Up verbs | A copy of the slide, emails |
| Tone updates “Tone & Clarity” | A verifying voice, few hedges | Board items, Priss summaries |
Reply to Universal Rewrite
Revise the paragraph for senior-executive tone; keep ≤120 words.
Retain numbers and units; add one persuasive stat if missing.7) End-to-end llm Camping
- The results of the model → Shape / metric → Description of data.
- EDA detection → Summary encourages or Langchain Chain.
- Self-check → Ask the model that a small size and unspecified or not in KPIS.
- Tone & Format Pass → re-submission to rewrite.
- Version control → store
.promptyfiles beside the brochures.
8) Subject lessons
| Org / Project | The use of the llm | Result |
|---|---|---|
| Functional credit card | Shape-to-Narrative (“Shapestories”) within Dashboard | + 20% understanding of stakeholders; 10 × Quick Docs |
| Health restart | ROC translator for a glittering app | Doctors match the 92% sensitivity in small minutes in minutes |
| Retail Analytics | Embedded table summed | Three hours writing is reduced to ~ 12 minutes |
| The Great Desk of Wealth | Research Q & A Helper | The Month Questions in 200k; ≈90% Satisfaction |
| World CMI group | A line of understanding with the llm | Quick Reporting of 30 District Market |
9) The best checklist
- Describe audiences, length, and tone in The first two lines All immediately.
- Feast Formal Input (JSON / tables) to reduce halucinations.
- Pour Evaluation (“Compliment quality 0-1;” The flag is lost KPI “).
- Bestow temperature ≤0.3 Finding decisive summaries; Suggested the boards of creative story.
- Never check the numbers Without units; Keep the real metrics visible.
- Control of the transformation + that comes out; tie Model types by audit routine.
10) Common issues and Guardrails
| Fright | Emblem | Commitment |
|---|---|---|
| Have invented drivers | Accounting claims are not blameless | Rigid White feature |
| Too much technical | The participants issued | Add “Grade 8th Learning Level” + Alooliy Business |
| Tone Mismatch | Slides / memos are not likewise | Run Batch Tone-Rewrite Pass |
| Hidden Roads | Make you miss small-n or samples | Force a Limitations Bullet in all times |
These “mirrors begin to” mirrors “indicate how I closed my DS-LifeCycle pieces, because misuse is likely happening early, at the time of going.
The steal-this – take place to replace: Manage all metric as a story to be told, and then use to make it possible. Keep actions closer to, lilies near, and your word in unalaranced with yours.
Thank you for reading!
👉 Get a full cheat sheet + Weekly updates to applicable AI instruments when you sign up for Sara's Ai Automation in Gaya – Helping technical experts work real work with AI, every week. You will receive access to AI Library.
I give up instruction in work growth and the revolution here.
If you want to support my workCan you buy me my own The coffee you like: Cappuccino. 😊
Progress
Improving a shaped rate translation using large language models
How to summarize the data table easily: Pute the embedded llm
Tell me a story! Xai driven by rationalizing with large modes of languages
Using llms to improve data connection – Data Data
