Agent Prompting Guidelines

SOP Guidelines

Executing SOPs

1. Use the format: “Run X SOP with Y”, where:
   1. X is the closest matching playbook name.
   2. Y is the variable the user needs to specify.
2. For better accuracy, reference the exact variable or SOP that requires substitution:
   1. For variables, you can use the format $Y.
   2. For SOP, you can use the format Run < X SOP > where X is the name of the SOP saved in the platform.
   3. The final name can "Run <X SOP> with $Y"
      Note: The <> and $ can be extremely useful when you're giving additional verbose instructions alongside the prompt. For example "Run <X SOP> with $Y in context of the error that I'm seeing today on the ELB".
3. Example:
   ✅ "Run the Pod Health SOP with namespace 'x'."
   ✅ "Run the Node Disk Cleanup SOP with node_name 'Y'."

Writing SOPs

1. Markdown Section:
   1. Clearly describe the problem the SOP addresses.
   2. Provide step-by-step instructions on resolving the issue.
   3. If the SOP executes a task, ensure the necessary variables are defined.
2. Chaining SOPs with Analysis:
   1. If an SOP needs to be followed by an analysis, phrase it as:
      ✅ "Run the Pod Health SOP with namespace 'X' and then give me an analysis of the pods with the most restarts."
      ✅ "Run the Kafka Consumer Lag SOP with cluster 'Y' and analyze topics with the highest lag."
   2. Using Transformer Functions:
      1. SOP outputs can be transformed into new variables for subsequent tasks.
      2. Example: Extracting failing pods from Pod Health SOP and using them in a Log Analysis SOP.

General Debugging & Investigations

1. Be Descriptive: Provide sufficient context for the agent about:
   1. The task you want to execute.
   2. The alert/issue you are investigating.
   3. The expected outcome.
2. Leveraging Integration Memory:
   1. Navigate to Integrations > Memory to add persistent context.
   2. Example:
      ✅ "Always check branch 'B' for repo 'X'."
      ✅ "For Kafka, prefer using 'consumer_group_status' for lag analysis."
3. Handling AI Hallucinations:
   1. We have spent a lot of efforts to avoid hallucinations or distractions for the agent, including but not limited to the chaos minimisation framework. If responses seem inaccurate, add more context to refine the agent’s understanding & report back with thumbs down 👎.

Continuous Improvement

1. Feedback is crucial:
   1. If the agent’s response is helpful, give a thumbs up 👍.
   2. If not, provide constructive feedback to improve the training process and give a thumbs down 👎.
2. Improve connector context: Leverage Memory
3. Improve company context: Leverage catalog to further improve the understanding of internal jargons and keywords