How to Use LLMs to Transform Data into Stories: A Practical Workflow

import pandas as pd
import json

# Example: load and aggregate a CSV
df = pd.read_csv('sales_2026.csv')
agg = df.groupby('region')['revenue'].sum().reset_index()
agg.to_json('aggregated_sales.json', orient='records')
print('Aggregated data saved to aggregated_sales.json')

You are an expert data analyst specializing in data storytelling.

**Context:** You have an aggregated sales dataset by region (JSON file).
**Objective:** Create a concise narrative (3 paragraphs) that highlights:
- The region with the best performance,
- The trend compared to the previous period,
- A key recommendation for the next quarter.

**Constraints:** Use simple language, include specific numbers, and end with a question that encourages the reader to reflect.

**Dataset:** {dataset_json}

**Output:**
1. Title
2. Subparagraphs (3)
3. Final recommendation (in bullet list format)

from langchain.chat_models import ChatOpenAI
from langchain.schema import HumanMessage

llm = ChatOpenAI(model='gpt-4o')
prompt = """[Full prompt as above]"""
msg = HumanMessage(content=prompt)
summary = llm.invoke([msg])
print(summary.content)

# Additional visualization suggestion
viz_prompt = f"Suggest a suitable chart to represent this information: {summary.content}"
viz = llm.invoke([HumanMessage(content=viz_prompt)])
print('Suggested chart:', viz.content)
"""
The result is a publishable story, accompanied by a visualization recommendation that you can create with tools like Matplotlib or Plotly.
5. Review and iterate with feedback
Use the LLM feedback loop to refine the tone, correct any inaccuracies, and ensure the narrative aligns with the audience. Tools like OpenAI Pando allow you to iterate quickly without downtime.
Effective prompts for data storytelling (practical examples)
Here are three reusable prompts that work with 2026 models:
Executive summary prompt
You are a senior data analyst. Take the following sales metrics dataset and create a 150-word executive summary that includes:
- Key KPIs (growth, market share)
- Three highlights
- One challenge and one recommendation.

Dataset: {dataset}

Output:
Scenario-based storytelling prompt
Write a narrative-style story that personifies our main product as a traveler. Describe its journey through 2026 adoption data, highlighting peaks, dips, and turning points.

Use an emotional tone, include concrete data, and end with an inspiring sentence.

Dataset: {dataset}

Output:
Comparative analysis prompt
Compare Q3 2026 performance with Q2 2026 for each region, highlighting:
- Percentage change
- A possible cause for each change
- A quick action suggestion.

Dataset: {dataset}

Output:
Current trends influencing data storytelling (2026)
This year’s innovations are redefining what LLMs can do with data:

• MoE inference speed
• OpenAI’s Git backend (Pando)
• Soniox STT v6
• Integrated multimodal visualization tools

Real-world use cases and applications

• Monthly sales reports
• Product performance monitoring
• Voice search data analysis
• Interactive investor presentations

Key takeaways and concrete actions

• Always define a clear narrative goal before inserting data into a prompt.
• Use a structured prompt with role, context, task, and constraints to achieve repeatable outputs.
• Leverage MoE optimizations to run large models locally without the need for external GPUs.
• Integrate Soniox STT v6 for seamless voice data pipelines.
• Try having the LLM generate both text and visualization code: it reduces steps and keeps the workflow consistent.
• Implement a rapid feedback loop (thanks to OpenAI’s Pando) to refine the narrative based on audience reactions.

Conclusion

In 2026, LLMs are no longer just text models: they are complete storytelling engines, capable of transforming raw datasets into engaging stories in seconds. By following the step-by-step workflow described above and using proven prompts, you can automate data storytelling, free up time for strategic analysis, and communicate insights more effectively than ever before.

Get started today: prepare a small dataset, experiment with a role-based prompt, and watch how LLMs transform numbers into narratives that your audience loves to hear.