Template Techniques

Progressive State Prompt Injection

Explains how to inject focused system instructions as an AI-assisted workflow moves between states.

A common technique on more complex business processes it to progresseively inject system prompts as the conversation progresses.
This is acheived in the Stubber Framework by injecting system prompts deterministically when certain actions are run OR when a state is entered.
This allows keeping the main system prompt high level and leaving detailed instructions for the state in which they will be needed.
It allows a token efficient way of building deterministic AI Agents

# Detailed Example Structure

• A support conversational agent begins a conversation with a customer
• The agent runs an action called verify_customer with the customer identifier
• This action transitions the stub state into "customer-support-in-progress"
• The "customer-support-in-progress" state has an "on-enter" gpt_chat_task task
• The task injects a system role message into the main chat to give the agent specific instructions about how to support the customer
• This keeps the details of how support the customer out of the context until the customer has been verified