The six-part prompting framework was where I started with AI. It gave me a way to understand what I was asking for and gain more control over the conversations I was having.
I still use it whenever I design a conversational agent: Persona, Context, Task, Format, Exemplars, and Tone.
My workflow has developed since then. For recurring tasks, I now save instructions in Markdown skill files and use scripts where the work needs executable logic. For a conversational agent, this framework remains my starting point.
The framework helps me design the instructions. A file gives those instructions a reusable home.
The 6 Components
1. Persona
Who is the AI in this interaction?
Define the role and its purpose. For example:
"You are a customer-support guide for an online store. Help customers understand the published delivery and returns policies."
The role should describe the behavior you need. Calling a model an expert does not give it credentials or make its answers automatically correct.
2. Context
What's the situation?
Give it the relevant background, sources, and boundaries:
"Use the approved policy documents provided with the conversation. Customers may not know their order status. Do not invent tracking information or policy exceptions."
Context is where I spend time. The agent needs to understand the situation it is supposed to serve, not just the topic it should discuss.
3. Task
What exactly should it do?
Be specific about the job:
"Answer the customer's question, identify the supporting policy, and ask for missing information when it is necessary to help."
This gives the conversation a purpose. It also helps distinguish answering a question from taking an action such as changing an order.
4. Format
How should the output be structured?
"Begin with a direct answer, follow with the relevant policy detail, and give one clear next step when there is one."
For a conversational agent, format is a response pattern rather than a requirement to produce the same rigid template on every turn.
5. Exemplars
What does a good answer look like?
An example makes the intended behavior concrete. This one is illustrative; it is not an actual store policy:
Customer: "Where is my package?"
Agent: "I don't have your tracking status in this conversation. Use the tracking link in your dispatch email, or contact support with your order reference so they can check."
Examples should match the sources and capabilities the real agent has. A sample answer should not teach it to promise an action it cannot perform.
6. Tone
How should it sound?
"Use plain language. Be warm, concise, and practical. Avoid blame, exaggerated enthusiasm, and technical jargon."
Tone matters because the same information can feel helpful or frustrating depending on how it is expressed.
The Missing Piece: Let the AI Interview You
Even a structured prompt can leave important gaps. I want the AI to ask useful questions when it needs my input rather than guess what I meant.
That is how I approached refreshing this blog. I asked the agent to interview me about the articles so it could capture my actual experience, examples, and perspective before revising them.
For an agent-design session, I might ask:
"Before drafting the instructions, ask me what you need to understand the agent's purpose, audience, sources, actions, and boundaries."
That does not mean every conversation should begin with a long questionnaire. Ask when the answer changes the design. Once there is enough context, produce something concrete that we can improve together.
The Full Prompt
Combining the six parts gives a clearer starting instruction:
"You are a customer-support guide for an online store. Use the approved delivery and returns policies supplied with the conversation. Answer the customer's question, identify the supporting policy, and give a practical next step. Ask for missing information only when it is needed to help. If the documents do not answer the question, say so and direct the customer to support. Do not invent order status, policy exceptions, or actions you have not performed. Begin with the direct answer and keep the explanation concise. For example, if tracking information is unavailable, explain where the customer can find it rather than inventing a delivery date. Use warm, plain language."
The actual sources, tools, and business rules still need to be connected and tested. This prompt defines the intended behavior; it does not implement the whole system.
From a Conversation to a Reusable File
When I am building a repeatable workflow, I do not want to reconstruct the same instructions in every thread. I capture the procedure in a skill file and use Python or other scripts where a step should execute as code.
A conversational prompt can also live in Markdown. The distinction is the purpose: the six-part framework shapes how an agent interacts, while a workflow skill describes a procedure it can reuse. A project can need both.
The Prompt Generator
My JedAI Prompt Generator helps apply this framework to a conversational agent.
Tell it what you want to create: the agent's purpose, who it should help, what information it should use, what you want included, and how you want it to respond. The more relevant context you provide, the better starting prompt it can produce around the six components.
That starting prompt is something to test and refine. For me, useful prompting has always been a conversation: explain the goal, let the agent identify gaps, try the result, and improve the instructions.