Introduction: Why data storytelling with LLMs is essential today
How to structure an effective data narrative with LLMs
Step 1: Transform data into a readable format
Start by converting your dataset into a compact JSON string or a markdown list. This ensures the model receives structured, easily analyzable information.
{
"dataset": [
{"month": "January", "sales": 12000, "region": "north"},
{"month": "February", "sales": 15000, "region": "south"}
]
}Step 2: Write a basic prompt
Use a clear, goal-oriented instruction. A universal template works for most use cases.
Generate a concise markdown story summarizing the following data: {data}. Highlight trends, anomalies, and suggestions. Use a tone suitable for non-technical executives.Step 3: Refine the prompt with constraints and context
Add constraints such as length, output format, and style. The more specific you are, the better the result.
Generate a story of maximum 300 words in Italian, based on the data provided. Include a 'key points' section with bullet lists and a conclusion with an actionable recommendation.Practical example: From Bluesight’s hospital dataset to a compliance dashboard
Bluesight developed Prism, an AI layer that connects hospital pharmacy data to 340B compliance. Using an LLM, Bluesight engineers created an assistant that transforms complex distribution logs into narrative reports.
Here’s a reusable prompt for similar scenarios:
Analyze the distribution data of {data} and produce a narrative that highlights any deviations from the 340B threshold. Provide an executive summary, a trend graph, and a recommendation for an audit.Real-world cases in 2026: robotics and web crawling
Mistral AI launched Robostral Navigate, an 8B model that enables robots to navigate complex environments using a single RGB camera. Prompt engineers trained the model to describe the surrounding world in natural language, creating a visual narrative for robots.
Here’s an example prompt for this use case:
Given the RGB camera frame data {frame_data}, generate a natural-language description of the robot’s immediate environment, highlighting obstacles, pathways, and recommended navigation actions.Meanwhile, AI agent crawlers now require explicit permission to access web pages. Here’s an effective prompt for generating a consent statement:
Generate a concise, GDPR-compliant consent notice in Italian that explains why this agent is accessing {url}, what data will be collected, and how it will be used, asking for user permission.Tools and resources to get started
- Prompt editors likeWeights & BiasesorPromptistfor quickly testing variants.
- Public datasets onKaggleandHugging Facefor practicing with realistic data.
- Predefined prompt blocks inOpenAI,AWS Bedrock, andGCP Vertex AIthat you can copy and adapt.
Common mistakes to avoid
- Prompts that are too vague:Without constraints, LLMs often produce generic output.
- Neglecting data format:Providing plain text instead of structured JSON reduces accuracy.
- Ignoring cultural context:A prompt in Italian may require a specific tone or regional references.
Conclusion: Transform data into stories that inspire action
Key points
- Adapt constraints (length, tone, format) for each audience.
- Explore real-world use cases like Bluesight’s Prism or Robostral Navigate for inspiration.
- Always test prompts with a small sample before scaling up.