14 · Prompt Engineering Architecture

This shows a prompt engineering architecture for enterprise agents: the right context plus instruction plus memory equals better answers. It details how a final prompt is assembled from many sources.

How it works

1. 1System instruction: role, rules, constraints and tone.
2. 2Short-term memory: recent conversation history.
3. 3Long-term/user memory: user facts, preferences and past interactions.
4. 4Retrieved context (RAG): top relevant chunks from the knowledge base.
5. 5User question: assembled into the final prompt sent to the LLM.

Key points

• Put the most important instructions in the system prompt.
• Keep short-term memory small and relevant; retrieve less, but highly relevant, context.
• Always provide citations and use structured output for reliability.
• Principles: be clear, provide context, use memory wisely, ground with data, structure output.

Why it matters

In enterprise agents, the prompt is assembled programmatically from many sources. Getting that assembly right — instruction, memory, context — is what produces accurate, grounded answers.

Frequently asked questions

What goes into a production prompt?

System instruction, short- and long-term memory, retrieved RAG context, and the user's question.

How do I keep prompts reliable?

Prioritise the system instruction, retrieve only relevant context, cite sources and enforce structured output.