System, Role and Contextual Prompts
The reply-drafting prompt had grown into one long paragraph: rules about length, a request to sound friendly, today's repair queue, the customer's service history and the email itself, all jumbled together. Every time the owner updated the queue length, a rule got accidentally deleted. Splitting the prompt into layers solved it.
The rule
System sets the rules, role sets the voice, context sets the facts.
- System prompt — standing instructions that apply to every conversation: boundaries, format rules, what never to do. Chat APIs usually have a dedicated system message for this, and models are generally trained to give it more weight than ordinary messages.
- Role prompt — who the model should act as: "You are an experienced bike mechanic who explains things simply." A role packs expertise, vocabulary and tone into one line.
- Contextual prompt — the facts for this task: the customer's history, current repair times, the relevant policy. It changes on every call.
[System]
You draft email replies for Spoke & Chain. Keep replies under 120 words.
Never promise refunds; say the team will review them. Sign off as
"The Spoke & Chain team".
[Role]
You are a friendly, experienced bike mechanic who explains things simply.
[Context]
Customer: Priya. Bike: 2019 city bike, serviced 2 March.
Current repair queue: 5 days.
[Email]
My gears are slipping again. How soon can you look at it?A good reply would mention a slot in about five days, explain simply, stay under 120 words and sign off correctly — each layer doing its own job.
The shortcut
For every sentence in a messy prompt, ask "Is this true for every call, or only this one?"
- True for every call, and a rule → system.
- True for every call, and about voice → role.
- True only now → context.
"Keep it under 120 words"
A rule for every reply. System.
A role improves tone and focus, but it does not add knowledge. "You are a world-class mechanic" will not make the model know Spoke & Chain's stock; that still belongs in context.
Drills
Answer fast — trust the rule.
Where should "Never promise refunds" go?
"This customer's bike was serviced on 2 March." Which layer is this?
What does a role prompt mainly add?
Why keep the three layers separate instead of one long paragraph?
Remember
- System sets the rules, role sets the voice, context sets the facts.
- Shortcut: "true for every call, or only this one?"
- Trap: expecting a role to supply knowledge it was never given.