Think Before You Prompt:
The SPACE™ Framework Exercise
Use case: Designing an AI assistant for a call center
Your product team is building an AI assistant for TeleServ, a large telecommunications company. The AI will support human call center agents — not replace them — by suggesting responses, surfacing policy, and summarizing call context in real time.
Customers calling in are often frustrated. Calls can involve billing disputes, outage complaints, upgrade requests, and service cancellations. Agents are measured on AHT (Average Handle Time), FCR (First Call Resolution), and CSAT (Customer Satisfaction Score).
Write a prompt for the TeleServ AI assistant — whatever comes to mind first. Don't overthink it. This is your gut instinct. You have 5 minutes.
"You are a helpful AI assistant for a call center. Help agents respond to customers."
Sound familiar? Most instinct prompts look like this. We'll find out why that's a problem.
- What assumptions did you make about who uses this AI?
- Did you define what the AI should not do?
- How would the AI know when to stay silent?
- What happens if a customer says they want to cancel service?
Use these five dimensions to architect your prompt deliberately. Each one closes a gap that instinct prompts leave open.
| S | Scope | What is the AI allowed to do? What is explicitly out of bounds? Define the domain boundary — and the edges. |
| P | Principals |
Principals are not just users — they form a hierarchy of authority over the AI's behavior. The AI takes direction from principals in order of rank: the organization's policy sits at the top, then the operator (the team deploying the AI), then the end user. Each level can grant or restrict what the level below it can do — but cannot exceed its own authority. Ask: Who instructs this AI? Who can override it? Who is affected by its outputs but has no direct control? In a call center: TeleServ policy sets the ceiling → the product team configures the system prompt within that ceiling → the agent interacts in real time → the customer is affected but gives the AI no direct instructions. Escalation pathways are a Principals question: when the agent's authority runs out, who or what does the AI defer to next? |
| A | Actions | What specific behaviors should the AI exhibit? Suggest language? Surface policy docs? Flag escalation risk? Summarize? |
| C | Constraints | What must never happen? Tone rules, compliance limits, data privacy, escalation triggers, honesty requirements. |
| E | Example | Give a concrete example of a good interaction. Show the AI what success looks like — don't just describe it. |
Use the SPACE framework to write a new, more intentional prompt for the same TeleServ AI assistant. Work through each dimension below first, then write your final prompt at the bottom.
Read both prompts aloud with a partner. What changed? What would have gone wrong?
Instinct Prompt — What's Missing?
- Who is the primary user of this AI?
- What happens when a customer is angry?
- Can the AI promise a refund?
- Does the AI know when to escalate?
- Is there any example of a good output?
- What does "helpful" actually mean here?
SPACE Prompt — What Improved?
- Role and scope are explicit
- Principals are named (agent, customer, TeleServ)
- Actions are specific and enumerable
- Constraints prevent harmful outputs
- An example grounds the AI's behavior
- Tone, escalation, and limits are defined
What was the hardest part of filling in SPACE? Why?
Constraints is usually hardest. People resist defining limits — they feel restrictive. But constraints are protective, not punitive.
Who defined principals? Did anyone include the customer as a principal?
Most instinct prompts center the agent and forget that customer trust is also at stake — and the AI affects it directly.
What would happen if a customer threatened to cancel — using your Round 1 prompt?
The instinct prompt gives no guidance. The AI might blurt a discount, say nothing, or contradict company policy.
What's the cost of a vague prompt in a high-stakes context like this?
Brand risk, compliance risk, agent distrust, customer harm. SPACE isn't just structure — it's protection.
Prompting without SPACE is guessing with high stakes. The framework isn't overhead — it's the thinking you were going to do anyway, made visible and rigorous.