How to Write a Prompt to Analyze and Classify Data in 2025 (Full Guide)

To write an AI prompt for data analysis or classification, define the decision, give explicit criteria, separate each source of data, and specify the output columns. Include a rule for missing or contradictory information so the model can return Unknown instead of forcing a guess.

I use a company-classification example in this guide: given a company description and scraped website text, return whether the business sells HR services to other companies, plus a short reason. That gives you a result you can filter and a reason you can check against the input.

A clearer prompt makes the results easier to review. Test for missing evidence and incorrect classifications. Test representative rows before running it across your dataset.

Here are the topics we will cover in this guide:

1. What is a Prompt?
2. Examples of a Prompt
3. Who This Guide Is For
4. Why Good Prompts Matter
5. How a Bad Prompt Looks Like
6. Prompt Breakdown and Implementation Steps

But What is a Prompt Anyway?

Simple answer: A prompt is the input you give to an AI to get a specific output.

Don't be confused if you have read other articles that wrote lengthy paragraphs about "what is a prompt," because in reality, a prompt is not complicated to define.

What's tricky is writing a good prompt, but before I show you how to write one, let me give you 3 examples of a prompt.

Who This Guide Is For

Writing effective prompts is a challenge that everyone will face at some point in their career.

That's why this guide isn't designed for a specific industry or role but rather for those working with data and automation — we'll specifically focus on writing prompts that help you edit, analyze, and clean large datasets efficiently, whether you're a student or experienced professional.

The examples use Datablist to apply a prompt to spreadsheet rows and write the results into separate properties.

Why Good Prompts Matter

A model needs the decision rules that a colleague would need to classify the same row. Clear instructions reduce ambiguity, but incomplete inputs and model mistakes still need a review step.

How a Bad Prompt Looks Like

Why This Prompt is Bad:

• No structured instructions: The requirements are scattered throughout the text without clear organization.
• No clear sections or separators: The AI has to parse through a wall of text without any logical breaks.
• Unlabeled inputs: The company description and website text have no clear boundaries, making their roles harder to distinguish.
• Informal language: Using phrases like "stuff" and "btw" creates ambiguity.

How the AI Sees the Prompt

Here is an illustration with the row values substituted directly. Datablist can rewrite inline variables into a separate Variables section; labels still help you identify each source.

Mixing input values into an informal instruction makes it harder to inspect where the task ends and the evidence begins.

Here's what goes wrong in this case:

1. Unclear input boundaries: The two sources run into the surrounding instructions.
2. Undefined labels: The prompt does not say what to return when evidence is missing or the company does several kinds of work.

👉 With that being said: let's now break down how a good prompt looks like and how to create one!

3 Examples of a Prompt

One Command Prompt

In its simplest form, a prompt can be a command like:

Question Prompt

A prompt can be a simple question like:

Structured classification prompt with tagged inputs

A prompt can also be a more detailed set of instructions that can edit, analyze, or clean your data to personalize cold emails, analyze customer feedback, clean company names, and more.

I use labeled sections to separate rules, output requirements, and input data. XML-style tags are one option, especially when a prompt combines several documents. Anthropic's prompting guide recommends descriptive tags for that separation. Plain section headings can also make a short prompt easy to inspect.

Here is a copyable classification prompt with an explicit third label for missing evidence:

For example, “We provide payroll software to employers” supports Yes. “We build accounting software and are hiring engineers” supports No. An empty description supports Unknown, even when the company name sounds related to HR. These are illustrative expected labels, not measured model results.

Prompt Breakdown and Implementation Steps

The screenshots below show the earlier template structure. Apply the same separation of task, criteria, and data to the updated classification prompt above, including its output fields and Unknown rule.

A Breakdown of Our Prompt Structure (The Good Prompts)

When you look at some of our AI prompts templates in the Datablist.com AI-prompts library, you'll notice that almost every prompt template is built out of sections and each section has these 3 components:

1. Tag: Lets the AI know what the section is about
2. Body: The body contains instructions that help the AI understand and complete the task through:
   • Explanation: Tells the AI something it should know/Gives the AI context
   • Command: Tells the AI what to do/Gives the AI a task
3. Boundaries: Headings, separators, or paired tags show where each section starts and ends. Use the same labels consistently; the delimiter cannot compensate for missing decision rules.

At the end we also have: Placeholder properties:

• Columns that get replaced with data when the prompt is used.
• These placeholders are used to customize the prompt for each record in a dataset.

For example, in our Job Title Categorizer prompt we use {{job_title}} as a placeholder that gets replaced with an actual job title from our dataset when processing each record.

The Breakdown of My Prompt & Implementations Steps

Note: HR Companies aren't our ICP ("Ideal Customer Profile") — we just used it as an example.

First Step: Context

When providing context to the AI, I did three key things:

• Provided information about what data the AI would receive, without revealing the specific targeting criteria
  • This explains the source inputs before the classification rules
• Explained the context around the data and outlined the specific problem
  • This helps the AI understand what it's working with
• Described my specific situation
  • This allows the AI to tailor its approach to my needs

What you have to do when writing your prompt:

• Tell the AI about the data you're going to give it
• Describe your problem
• Use separators (===)

Second Step: Task/Command (What I Want You To Do)

In this section of the prompt, I did two crucial things:

• Framed the AI again with the tag "What I want you to do" so it knows what I am going to talk about
• I gave it just the broad description of the task without mentioning the steps to take or mentioning which data to look for, to keep the information flow logical

What you have to do when writing your prompt:

• First, identify the core purpose of your task/command.
• Second, use separators (===)

If you have similar tasks to fulfill, don't try to do them all with one prompt as it would be very difficult for the AI to handle and for you to detect potential errors.

Third Step: Command Instructions (How To Do It)

These are the things I do when adding instructions to a prompt:

• Keeping it simple and straightforward
• Explaining the task as if I would teaching someone to do it manually
• Focusing on clear, step-by-step instructions

If the results are wrong, check the input text, decision criteria, and output requirements before rewriting the prompt. Some failures come from missing evidence or model limitations; adding more instructions cannot supply an absent fact.

What you have to do when writing your prompt:

• Create a simple step-by-step instruction using simple and clear language
• Use separators (===)

Fourth Step: Mistake Prevention no.1 (Context About the Task) – Optional

This step gives additional task-specific context to help the AI:

• Understand nuances of the task
• Avoid common pitfalls
• Make better decisions

By being explicit about potential issues and edge cases, we create guardrails that prevent the AI from making assumptions or mistakes.

This step isn’t required in every prompt but it is crucial for complex tasks.

What you have to do when writing your prompt:

• Think of common mistakes that junior employees make when doing this type of task
• Include these potential mistakes in your prompt to help the AI avoid them
• Use separators (===)

Fifth Step: Goal (What You’re Looking For)

These are the most important things to include when writing your expectations in a prompt:

• Providing detailed context about the desired outcome
• Being specific about expected results
• Avoiding revealing biases or desired conclusions that could influence the analysis

Using natural, conversational language instead of technical jargon typically leads to better results from AI systems.

What you have to do when writing your prompt:

• Define clear outcomes you want to achieve from the analysis or task
• Be specific about what indicators or signals matter most to you
• Give it examples
• Use separators (===)

Sixth Step: Data Context (Mention About the Data) – Optional

When I add data context to my prompts, I follow these rules:

• Reinforce mistake prevention through clear rules and gave it a reason
  • Rule example: "Always evaluate complete context"
  • Reason: "Single indicators can be misleading"
• Add multiple layers of context
  • Potential mistakes to avoid
  • Data-specific issues
  • Task-specific reminders

What you have to do when writing your prompt:

• Think of potential mistakes the AI could make and explicitly instruct it to avoid those specific errors.
• Split information into logical parts instead of providing it all at once.
• Use separators (===)

📘 Quick Fact

By being explicit rather than letting the AI make assumptions about the prompt, you will prevent many mistakes.

Seventh Part: How to Start (First Step)

This final step involves providing clear starting instructions to the AI. Here's what I did:

• Kept it simple and straightforward
  • Numbered the steps chronologically
  • Used clear action words like “read, make, decide…”

What you have to do when writing your prompt:

• Give clear starting instructions with numbered steps
• Use action verbs to make the steps easy to follow
• Keep instructions simple and straightforward
• Use separators (===)

Eighth Step: Placeholder Format (Always Last)

This is the final step in writing an effective prompt: insert columns as a placeholder.

Keep stable task instructions together, then put row-specific inputs in labeled sections at the end of this template. That makes the prompt easier to maintain and inspect; input order alone does not ensure accuracy.

What you have to do:

Use double curly braces {{ }} to add data from your spreadsheet columns

Use separators (===)

❗Important

Select the actual collection properties when inserting variables. Keep their values in labeled input sections and review the generated output before processing the full list.

Rules Applied on Another Example

The following M&A scenario illustrates how to organize a more complex request. It needs precise criteria before it can be used for company scoring.

Example request:

Let’s start prompting!

Step One

In this step, we'll give the AI general context.

Step Two

Now we will clearly outline the main objectives for our AI assistant.

Step Three

Here we'll break down the specific steps for our AI to follow in its analysis.

Step Four

To ensure accuracy, we'll now explain potential pitfalls and important considerations.

Step Five

Now we’ll define what success looks like by establishing clear evaluation criteria.

Step Six

Now we'll provide context about data interpretation and industry specifics.

Step Seven

Now we’ll tell the AI the exact sequence of actions to begin its analysis.

Step Eight

And lastly, we'll give the AI data to work with.

Adapt the criteria and output fields to your task. The M&A example still needs explicit scoring thresholds and a human review; a general request to assess financial health is not a defined scoring model.

Test the prompt before processing your CSV

For the HR example, enable Define outputs format and configure HR Match and Reason as Text outputs in Datablist's Ask ChatGPT/OpenAI enrichment. Map each output to its own property. A Checkbox would lose the distinction between No and Unknown.

Start with a small selection that includes an HR provider, an unrelated company that mentions hiring, an empty description, a mixed-services company, and conflicting source texts. Write the expected label for each row before running the prompt, then compare both the returned label and its reason with the original text.

When a result differs, identify the cause:

• A missing source value calls for better input data or an Unknown result.
• An ambiguous business description calls for a more specific inclusion or exclusion rule.
• A reason unsupported by the input calls for a correction and another test.
• An unexpected label calls for clearer output rules or output configuration.

Retest the same cases after revising the prompt. Keep a few additional rows for a separate check so you do not tune only to the examples already reviewed. All-empty variables may be skipped with an error rather than sent to the model, so inspect run errors as well as classification outputs.

To apply the checked prompt to a larger file, follow the CSV row enrichment workflow. Keep the original input properties for later review.

Conclusion

Start with one decision, explicit criteria, and separate output fields. Keep source texts labeled, allow a result for missing evidence, and test the prompt on ordinary and difficult rows before processing the full collection.

Frequently Asked Questions About Writing AI Prompts

What is a good AI prompt?

A good AI prompt is clear, specific, and well-structured. It includes context about what you want to achieve, specific instructions on how to do it, and the desired format for the output. The prompt should be written in simple language and avoid any ambiguity that could lead to misunderstandings.

Check our AI prompts examples for lead generation.

Why are my ChatGPT prompts not working?

AI prompts might not work effectively because of:

• Unclear or vague instructions
• Too many tasks in a single prompt
• Missing context or background information
• Complex or technical language
• Lack of specific examples or desired output format

How do I write AI prompts for better results?

To write effective AI prompts:

• Start with a clear objective
• Break complex tasks into smaller steps
• Provide relevant context and background
• Specify the desired output format
• Use simple, straightforward language
• Include examples when possible

What are the most common mistakes in prompt writing?

Common prompt writing mistakes include:

• Being too vague or general
• Overloading the prompt with multiple tasks
• Not providing enough context
• Using unclear or ambiguous language
• Assuming the AI understands implied information
• Not specifying the desired output format

How long should an AI prompt be?

An AI prompt should be as long as necessary to convey all essential information, but not longer. Typically, effective prompts range from a few sentences to a paragraph for simple tasks, and may be longer for complex tasks requiring detailed instructions and context. The key is to be comprehensive yet concise.

What is prompt engineering?

Prompt engineering is the practice of designing and optimizing inputs to AI models to get the most accurate and useful outputs. It involves understanding how AI models interpret instructions, structuring prompts effectively, and iteratively improving prompts based on results.