Explain your Goal to Bob

To get started, simply tell Bob what you need. For example, you could tell Bob, "I need an Agent that will help me analyze our paid search campaigns."

Bob will likely need more details on what types of insights are important to you or how you measure success in a campaign. You may say something like, "Great question! We typically optimize for a low cost per conversion, aiming to keep it below $5. I'd like to see the Cost per Conversion for all currently active campaigns and get recommendations on how to optimize them."

With this information, Bob will configure the Agent and, within seconds, complete the General tab—including a well-structured prompt ready for your review.

Understanding the Prompt Bob Writes

Think of the prompt Bob writes like instructions to a new team member—the clearer, more specific, and more actionable the instructions, the better the results. Bob is trained to write prompts in this format, so you'll recognize the structure when reviewing his drafts.

Key Elements of a Good Prompt

A strong prompt includes four parts. When reviewing what Bob has written, here's what to look for in each:

1. Objective

The objective defines what the Agent should accomplish. When reviewing Bob's draft, check that this is concrete and specific—not vague. A good objective tells the Agent exactly what task it's performing.

• Example: "Objective: Benchmarking. Compare key metrics of a given advertiser against the average of those same metrics across all accounts in the same industry."

2. Process

The process breaks the task into organized steps that guide the Agent toward the correct output. When reviewing Bob's draft, check that the steps are clear and sequential. The more complex the task or the more datasets involved, the more important it is to have well-defined step-by-step instructions—just as you would when training a human.

• Example:
  • Process:
    • Step 1: Locate the given advertiser name in the Google Ads dataset. If the user has not given you an advertiser name, ask for one.
    • Step 2. Extract Cost and Clicks for the given individual Advertiser for the last 30 days. If that advertiser does not have data in the last 30 days, inform the user of that.
    • Step 3. Then, identify all accounts that are associated with the same industry group that the individual account is associated with, using the Account_Groups dataset.
    • Step 4. Extract Cost and Clicks for all those accounts that have data in the last 30 days. Not all of these accounts may have data in that timeframe. That's ok.
    • Step 5. Calculate the group average Cost per Click (Sum of Cost for all accounts DIVIDED BY Sum of Clicks for all accounts)
    • Step 6. Compare the individual account's performance against the group benchmark. Present your findings in a table format where there are 2 rows: 1. [Individual Account Name] 2. [Industry Account Group Name Average] and present your calculations for each. Formatted table example below:
      | Comparison | CPC |
      |---------------------------|-----------------------|
      | [Individual Account Name] | $3.50 |
      | [Industry Account Group Name] | $2.80 |

3. Important Notes

This section captures guidelines the Agent should keep in mind while completing the task. When reviewing Bob's draft, check that any critical constraints—like time windows, data handling rules, or output standards—are captured here. Repetition helps AI—just like it does with humans.

• Example:
  • Important:
    • Make sure you're always analyzing the last 30 days worth of data, unless otherwise instructed by the user.
    • Check your work before presenting the results to the user.

4. Sample Query (Optional)

For tasks requiring precise data retrieval, Bob may include an example SQL query to help the Agent understand how to extract relevant information. You can also embed SQL directly in the prompt or instructions when needed.

• Example:
• Query to Execute:
  SELECT "Cost", "Clicks"
  FROM "Google Ads-Campaign Performance"
  WHERE "Advertiser Name" ILIKE '%Nike%';

Tips for Refining Prompts

Bob's first draft is a strong starting point—but refining it is where you make the Agent truly yours. Here are four tips to guide that process:

• Be specific. When reviewing Bob's draft, check that the objective and process are concrete, not vague. Instead of a general instruction like "analyze data," the prompt should specify what data to analyze, how to analyze it, and what kinds of insights to surface in the output.
• Anticipate outputs. Make sure Bob's prompt describes the output you actually want—whether that's a downloadable CSV, a chart, or a bulleted list formatted for easy copying into an email.
• Start simple. Don't overload Bob's first draft. Get the core task working before layering in complexity. Test that the results are correct and in the desired format, then build from there.
• Iterate and test. Run the agent, see where it falls short, and refine the prompt with Bob. Crafting prompts is an ongoing process—each round of feedback makes the Agent more reliable.

Find Relevant Datasets

Next, Bob will scan the Datasets you've already imported into Data Cloud and surface the ones most relevant to your request.

Continuing with the example above, since you mentioned "paid search" campaigns, Bob might identify datasets from Google Ads and Microsoft/Bing that contain the necessary fields for optimizing Cost per Click (CPC). He'll list these for you and ask which ones you'd like to assign to the Agent.

• If Bob suggests the correct datasets: Confirm the selection, and he'll assign them to the Agent.
• If Bob doesn't find the right datasets: You can manually tell him which ones to use or assign them yourself in the Knowledge tab.
• If Bob still can't find the datasets: Check whether they've been added to Data Cloud. If they're missing, you'll need to import them first. Refer to How to build a Native Provider Dataset in Data Cloud, then return to Bob once your data is in place.

Manage Knowledge Files

Bob can also create and manage Knowledge Files for your Agent. While building or editing an Agent, Bob will determine what information belongs in the Agent's core instructions versus what should live in a separate Knowledge File.

Here's how Bob thinks about it:

• Agent Instructions should cover the core stuff the Agent needs every time — its role, tone, and key rules.
• Knowledge Files are better for reference docs, FAQs, process guides, and other supplementary context — things the Agent can look things up when they're relevant, without being overloaded upfront.

This means you get a better-organized Agent with better-managed context, all handled by Bob.

What Bob Can Do with Knowledge Files

Bob can create, read, update, and delete Knowledge Files that he has generated. When Bob creates a file, it appears in the Knowledge tab of the Agent Builder just like a file you'd upload yourself.

Supported file formats for Bob-generated files include: .csv, .txt, .md, .json, .py, and other text-based formats.

📘

Bob-generated vs. User-uploaded files

Bob can only edit or delete files that he created. Files that you upload yourself are protected — Bob cannot modify or remove them. This ensures your manually curated content stays intact.

How It Works

1. During a conversation with Bob, describe what your Agent needs to do. Bob will determine if any supplementary context should be organized into Knowledge Files.
2. Bob will surface the file he's created in the conversation and ask for your confirmation before assigning it to the Agent.
3. Review the file — Bob will show you the content he's written. You can ask him to make changes or approve it as-is.
4. Once confirmed, the file is assigned to the Knowledge tab of the Agent automatically.
5. If you need changes later, you can ask Bob to update or delete files he created — just ask him in conversation.

📘

You can also always manage Knowledge Files manually in the Knowledge tab of the Agent Builder, regardless of whether they were created by Bob or uploaded by you.

External Resources

AI prompt engineering is evolving rapidly, with new strategies emerging regularly (e.g., zero-shot, few-shot, chain-of-thought). Staying up to date with best practices can help you build even more effective AI Agents in NinjaCat. Here are some great resources to explore:

OpenAI Prompt Engineering

Prompt Engineering Guide