About prompting Atlas AI agents

Atlas AI agents answer questions about your industrial data by querying the CDF knowledge graph. The quality of your prompts directly affects the accuracy and usefulness of agent responses. Specific prompts that include context and define what you need help the agent retrieve the right data and return a useful answer the first time.

Be specific and clear

Specific prompts produce better results than vague ones. Use action verbs to clarify what you want the agent to do.

Vague
Specific
Highly specific

Tell me about Pump P-101.

This prompt does not give the agent enough information to focus the response. The agent cannot determine whether you need operational status, maintenance history, specifications, or location information, which can result in an incomplete or unfocused answer.

What is the current operational status of Pump P-101?

By requesting operational status directly, this prompt helps the agent retrieve the right data and return the specific information you need.

List all open high-priority work orders for Compressor C-205 created in the last 30 days.

By specifying status, priority, equipment, and time frame, this prompt helps the agent filter and return only the actionable items you need.

Provide context

Context helps agents focus on relevant data by narrowing the scope of your request.

Without context
With basic context
With detailed context

Show me the temperature trends.

Without knowing which equipment, sensor, or time period to analyze, this prompt leaves the agent to search through data that may not be relevant. You may receive temperature trends for the wrong equipment or an unhelpful time range.

Show me the temperature trend for Reactor R-301, specifically for sensor TI-301A, over the past 7 days.

By specifying the equipment, sensor, and time frame, this prompt helps the agent retrieve exactly what you need.

Considering the recent maintenance on Heat Exchanger HE-502, are there any anomalous pressure
readings in the connected lines from the past 48 hours?

The maintenance context in this prompt helps the agent distinguish between expected post-maintenance behavior and readings that require attention.

Define the desired output

Specify the format, structure, and level of detail you need in the response.

Undefined
Defined format
Defined precision
Structured output

Give me maintenance data for Valve V-789.

This prompt does not tell the agent whether you need maintenance history, upcoming schedules, or cost data. It also does not specify a time frame or output format, which can lead to an unfocused or incomplete response.

Summarize the maintenance history for Control Valve CV-789 for the last 12 months, focusing
on repair actions taken and parts replaced. Present it as a bulleted list.

This prompt specifies the data type, time frame, focus, and format, giving the agent everything it needs to produce a useful response.

What were the maximum and minimum vibration levels for Motor M-607 on January 15th, 2024?

By requesting specific metrics rather than general information, this prompt tells the agent exactly which values to retrieve.

List all work orders for Pump P-101 in a table showing work order ID, status, description,
and date.

Requesting a table with defined columns makes the response easy to scan and compare across multiple items.

Use schema awareness

When your agent uses the Query tool, it can discover the structure of your data model at runtime. You can ask questions that rely on relationships between assets, equipment, and data without knowing the underlying field names or schema. This table shows examples of questions that use schema awareness effectively.

Question	What the agent does
How many maintenance orders are there per equipment type?	Queries the knowledge graph and groups results by equipment type
What sensors are connected to Compressor C-205, and which ones have readings outside the expected range?	Traverses the relationship between the asset and its linked time series, then checks current values
Show me all heat exchangers with overdue inspections	Filters on asset type and inspection status without requiring you to specify field names

The more clearly your organization’s data model is documented in CDF, the more accurately the agent can answer these types of questions.

Ask effective follow-up questions

Agents retain context within a conversation, so you can build on a previous response without repeating the original question. If the agent returned a list of work orders, you can follow up with “of those, which ones are high priority?” or “show those results as a table” without re-specifying the equipment or time range. Follow-up questions work well when they narrow, reformat, or extend the previous result. If you need to ask about a different piece of equipment or a completely different workflow, starting a new conversation resets the agent’s context. When a response is not what you expected, check the reasoning field to see how the agent interpreted your question and which data it used. If the agent queried the wrong equipment or applied the wrong filter, use that information to refine your next prompt with more specific criteria.

​Be specific and clear

​Provide context

​Define the desired output

​Use schema awareness

​Ask effective follow-up questions

​Related information

Be specific and clear

Provide context

Define the desired output

Use schema awareness

Ask effective follow-up questions

Related information