Template Techniques

Split-Brain Process

Explains how separate LLM contexts can divide responsibilities and reduce prompt-injection and persuasion risks.

Some processes require that a different LLM context process different parts of the process to ensure that prompt injection and persuasion techniques do not work.
A strategy that can then be used is to split the functions on the template between two chats having independent system prompts.
A conversational agent could pass details that require checking to a supervisor agent for example and then continue once the supervisor has given the go ahead.

# Detailed Example Structure

• A conversational agent is interacting with a user using the chat name main
• In one of the states there is an action called submit_for_approval which takes discrete fields for data used by the supervisor agent
• The submit_for_approval action starts up a new chat called approval with a system prompt for the supervisor
• The fields data gets substituted into the prompt for the supervisor
• The action moves the process into a state called "approval-pending"
• There are two AI enabled actions with the allowed_chat_names set only to the approval chat
• The two actions are approve and reject
• The approve action moves the state to "approved" and has a gpt_chat_task that injects a system role message into the main chat to notify of the successful approval
• The reject action moves the state back to the previous state and also injects the reject reason
• Both the approve and reject actions will have disabled_dynamic_tasks set to true, as the approval chat agent won't need any feedback after calling approve or reject