How Do LLMs Turn Data into Stories? Practical Prompt Examples

Introduction: Why LLMs are the engine behind modern data storytelling

How LLMs transform data into narrative: a step-by-step workflow

A typical data storytelling pipeline with an LLM follows five key steps:

• Data preparation:Upload your dataset (CSV, Excel, database) and generate a summary card with shape, columns, and basic statistics.
• AI-driven analysis:Ask the LLM to identify trends, anomalies, and correlations in the data.
• Story selection:Choose the most suitable narrative type (growth trends, segment comparisons, forecasts, or root cause analysis).
• Prompt writing:Provide the model with context, desired length, and style (formal, technical, or executive summary).
• Review and visualization:Refine the output, add charts or interactive embeds using tools like Streamlit or Power BI.

Each of these steps can be automated or semi-automated with just a few lines of code.

Practical example 1: Generate a story with a single prompt

Imagine you have a file namedsales_2026.csvwith columnsdate,region, andsales. You want a monthly executive summary-style story.

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

llm = ChatOpenAI(model_name="gpt-4o", temperature=0.2)

prompt = """
Analyze the following monthly sales dataset and generate a 150-word executive summary.
Highlight growth trends, seasonal peaks, and underperforming regions.
Format: introductory paragraph, key points, conclusion.

Dataset:
{dataframe.to_string()}
"""

message = HumanMessage(content=prompt)
story = llm.predict([message])
print(story)

The result is a coherent narrative that can be directly embedded into a slide or dashboard.

Practical example 2: Connect LLMs to Agensh for multi-agent storytelling

Microsoft has open-sourcedAgensh, a platform that orchestrates up to 1,024 autonomous LLM agents. With Agensh, you can delegate data collection, visualization, and writing to specialized micro-agents.

import agensh

# Define three agents
analyst = agensh.Agent(role="Data Analyst", tools=[load_csv, compute_stats])
visualizer = agensh.Agent(role="Graph", tools=[plot_bar, plot_line])
storyteller = agensh.Agent(role="Narrator", tools=[generate_text])

# Workflow
workflow = agensh.Workflow([analyst, visualizer, storyteller])
result = workflow.run(dataset="sales_2026.csv", question="Show sales trends by region in Q3 2026.")
print(result.story)

This approach is ideal for large enterprise reports where scalability and task delegation are critical.

Practical example 3: Optimize prompts with TokenGate for extended contexts

TokenGate, developed by Stanford and UC Berkeley, reduces prefill time by up to 11.3x for long contexts. Integrating it with your LLMs lets you handle datasets with thousands of rows without sacrificing speed.

# Example using Hugging Face Transformers and TokenGate
from transformers import AutoModelForCausalLM, AutoTokenizer
from token_gate import TokenGatePruner

model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-70b")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-70b")
pruner = TokenGatePruner(model, threshold=0.1)

inputs = tokenizer("Analyze this dataset:\n" + df.to_string(index=False), return_tensors="pt")
# Apply dynamic pruning
inputs = pruner(inputs)
outputs = model.generate(**inputs, max_length=500)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Result: Faster story generation without compromising quality.

Tools and resources: Anthropic’s OSS Scanner and beyond

Anthropic has launchedOSS Scanner, a free vulnerability scanner for open-source projects. While focused on security, you can use it to verify that the LLM libraries you rely on for data storytelling are up-to-date and secure.

• Integrate OSS Scanner into CI/CD pipelines to alert you about potentially risky dependencies.
• Use OSS Scanner’s security features to protect your datasets before analysis.

Practical tips and takeaways

• Start simple:A single prompt can already deliver a basic story. Add complexity only if needed.
• Leverage agents:For repetitive reports, Agensh lets you create reusable pipelines.
• Optimize with TokenGate:If you work with large datasets, integrate pruning to keep latencies low.
• Validate with OSS Scanner:Ensure LLM libraries are secure before deploying to production.
• Document prompts:Use README files or Jupyter notebooks to version your data storytelling prompts.

Conclusions

LLMs have made data storytelling accessible to everyone, transforming raw data into clear, actionable narratives in seconds. By experimenting with practical prompt examples, leveraging tools like Agensh, and optimizing performance with TokenGate, you can create scalable storytelling workflows that meet the demands of 2026. Remember to keep your tools secure with OSS Scanner and always document your processes for effective collaboration.