Template Techniques

Agent Guidelines

Explains how to expose focused guidelines to an AI agent on demand while keeping its main prompt concise.

In order to preserve tokens and break up a large number of specific instructions a good technique is to use guidelines.
Individual mini-prompts are defined under the data path prompts.guidelines, eg. prompts.guidelines.cancellations.
The AI agent is given an action to read_guideline which returns the respective guideline content.

# Detailed Example Structure

• Multiple guideline prompts are defined under prompts.guidelines.[guideline_name]
• A summary of the guidelines is defined under prompts.guidelines_summary
• An action called read_guideline is made available to the main chat in the _all state
• The read_guideline action description is Returns additional guideline content based on the guideline selected. {{{{skip}}}}{{{template.data.prompts.guidelines_summary}}}{{{{/skip}}}}
• This injects the summary of the guideline options into the action (and tool) description
• On the _create action a code task inspects the template.data.prompts.guidelines object and builds an array of objects to be used for a select field on the read_guideline action, is saves them on the stub as guideline_options
• The read_guideline action has a select field called selected_guideline whose options param is set to ~~stub.data.guideline_options
• When the AI agent selects a specific guideline that guideline is return in the ai_inject_return_data