Your First AI Prompts as a Support Agent
The quality of your AI output depends almost entirely on the quality of your prompt. Bad prompt, bad reply. Good prompt, and you'll wonder how you ever worked without AI. In this lesson you'll write your first support-focused prompts and learn the CRAFT framework that makes them consistently useful.
What You'll Learn
- The CRAFT prompt framework for support teams
- How to turn a vague prompt into a precise one
- Four ready-to-use prompts for daily ticket work
- Why context blocks beat clever wording every time
Why Prompt Quality Matters More in Support Than Most Jobs
When a marketer uses a weak prompt, they get mediocre copy. When a support agent uses a weak prompt, they might send a customer the wrong refund amount, cite the wrong return policy, or come across as cold to someone already upset.
Support is high-stakes because your output lands directly in a customer's inbox, often when they're already frustrated. So your prompting habits need to be more deliberate than the average ChatGPT user.
The CRAFT Framework
CRAFT is a five-part prompt structure built specifically for professional work:
- Context: What's the situation? Who is the customer? What product?
- Role: Who should the AI act as? (e.g., "You are a senior support agent for a SaaS company")
- Action: What exactly do you want it to produce?
- Format: How should the output look? (email, bullet points, under X words)
- Tone: What voice should it use? (empathetic, firm, cheerful, formal)
Let's see it in practice.
Bad Prompt
Write a reply to a customer who wants a refund.
This is garbage in, garbage out. The AI will produce something generic that may contradict your policy, pick the wrong tone, or miss the actual issue.
CRAFT Prompt
Context: A customer ordered our $89 noise-cancelling headphones three weeks ago. They emailed saying the right ear cup crackles when they turn their head. They've tried a factory reset. They're asking for a refund. Our policy allows returns within 30 days for defects.
Role: You are a senior support agent for a consumer electronics brand known for friendly, no-hassle returns.
Action: Draft a reply that apologizes, confirms the product is defective, offers a prepaid return label, and explains the refund will hit their card 3-5 days after we receive the headphones.
Format: Email, under 150 words, with a clear subject line.
Tone: Warm, apologetic but not groveling. Confident that we'll fix this.
Paste that into ChatGPT or Claude and you'll get a reply you can edit and send in under a minute.
Four Prompts You Can Use Today
1. The "Rewrite This More Kindly" Prompt
Paste your draft reply plus:
Rewrite this reply to sound more empathetic and less defensive. Keep the facts exactly the same. Target length: under 120 words. Our brand tone is warm and professional.
Great for those drafts you type when you're frustrated with a customer and need the AI to soften it before it goes out.
2. The "Summarize This Ticket" Prompt
Paste the entire ticket thread plus:
Summarize this ticket in 5 bullet points: (1) who the customer is, (2) what the issue is, (3) what's been tried already, (4) the customer's current emotional state, (5) the recommended next step. Keep each bullet under 20 words.
Perfect for handoffs between shifts or when escalating to a specialist.
3. The "Tag This Ticket" Prompt
Here is an incoming support email:
[paste email]
Classify it into exactly one of these categories: Billing, Bug Report, Feature Request, Account Access, Shipping, General Question. Also rate urgency (Low / Medium / High) and churn risk (Low / Medium / High). Respond in JSON.
Useful for team leads who triage manually before assigning.
4. The "Translate My Reply" Prompt
Translate the following support reply into natural, professional Spanish. Keep our friendly brand tone. Keep product names and URLs in English.
[paste your English reply]
Works for any language -- Claude and ChatGPT handle major languages well. Always have a native speaker on your team spot-check critical replies.
Context Blocks Beat Clever Wording
The single biggest upgrade you can make to your prompts isn't using fancier language -- it's pasting more context. AI can only work with what you give it. If you don't paste the customer's order number, product name, prior replies, and your policy, the AI has to guess, and guessing is where hallucinations happen.
A practical habit: before drafting a prompt, copy-paste these into a "context block" at the top:
- The customer's latest message
- The last 1-2 of your previous replies
- The relevant policy ("We offer 30-day returns on electronics")
- The product or plan details
Then ask for the specific output you want below that block.
Common Prompt Mistakes in Support
Mistake 1: Asking for too much at once. Don't ask the AI to categorize, reply, and suggest a knowledge base update in one shot. Break it into separate prompts.
Mistake 2: Not specifying length. Without a word limit, AI writes long. Customers hate long. Always set a cap: "under 100 words."
Mistake 3: Forgetting the audience. The way you write to a confused grandparent is different from a technical developer. Tell the AI who's on the other end.
Mistake 4: Trusting invented policies. If you didn't mention your return policy, AI will invent one. Always paste actual policy text when it matters.
Try It Now
Open ChatGPT or Claude and paste this prompt (using your own details):
Context: I work at [your company type]. A customer emailed saying [their issue]. Our policy is [policy].
Role: You are a senior support agent.
Action: Draft a reply that [what you want it to do].
Format: Email, under 120 words.
Tone: Warm, professional.
Compare that output to what you'd write cold. Most agents find the AI version is 80% ready after one prompt and 100% ready after a 30-second edit.
Key Takeaways
- Use the CRAFT framework: Context, Role, Action, Format, Tone
- Context blocks (customer message + policy + history) beat clever prompting
- Always specify length -- AI writes long by default
- Use the 4 core prompts: rewrite kindly, summarize, tag, translate
- Never assume AI knows your policies -- paste them in