Best Claude Prompts for Customer Service

Studies on AI in customer support consistently show something counterintuitive: teams that give agents AI-drafted responses verbatim tend to get worse CSAT scores than teams that use AI to generate a first draft the agent then personalizes. The reason is simple -- customers can tell when a response sounds canned, and a canned AI response often feels worse than a canned human-written template because it has a specific kind of fluency that reads as impersonal at scale.

The fix isn't better AI writing. It's better prompts that produce responses actually worth personalizing.

Before: The Weak Prompt

Write a response to an angry customer who wants a refund.

This produces a response that acknowledges the frustration, apologizes generically, explains the refund policy vaguely, and ends with "please don't hesitate to reach out." It could apply to literally any angry customer at any company. The customer will read it and feel processed, not helped.

Why It Fails

The prompt has no context, no constraints, and no definition of what "handled well" means for this specific situation. Without knowing the customer's actual complaint, the product, the company's actual refund policy, or the tone the brand uses, Claude defaults to averaging across thousands of customer service interactions it's seen -- which produces the median customer service response, not a good one.

There's also a specificity problem on the emotional side. "Angry customer wanting a refund" covers an enormous range of situations: the customer who received a broken product on a first order, the customer who's been a subscriber for three years and had one bad experience, and the customer who is genuinely trying to get a refund for something they clearly misused. Each of these calls for a meaningfully different response.

After: The Improved Prompt

Write a customer service response for this situation:

Customer message: [paste actual message]
Product/service involved: [product name, brief description]
Customer history: [e.g., "first order, arrived damaged" or "3-year subscriber, first complaint"]
Our actual refund policy: [paste relevant policy language]
Desired outcome: [e.g., "approve refund and retain the customer" or "deny refund per policy but keep goodwill"]

Tone: [your brand voice -- e.g., "warm and direct, never formal or corporate"]
Length: under 150 words

Don't invent policy details I haven't given you. If our policy doesn't cover this scenario, flag it rather than guessing.

What this does: Pasting the actual customer message and actual policy language removes the two biggest gaps that cause generic AI responses -- the prompt now has enough specificity to generate something a real agent could send with minimal editing.

Breaking Down Each Element

The "customer history" line changes the response more than most people expect. A three-year subscriber who's never complained before gets acknowledged differently than a new customer -- not because the policy changes, but because the relationship context changes the right emotional register. Claude picks this up clearly when you state it directly.

The "desired outcome" field is the one most CS teams skip, and it's usually the most important. A response aimed at approving a refund and retaining the customer looks completely different from one aimed at denying the refund while preserving goodwill -- but the generic prompt can't distinguish between them.

⚡ Pro tip: For high-volume ticket types (shipping delays, login issues, billing questions), build one dialed-in prompt per type rather than using a single generic prompt for everything. A shipping-delay prompt with your actual carrier SLA language and compensation policy produces responses that require nearly zero agent editing, while a generic refund prompt requires heavy revision every time.

Three Real Scenarios

A SaaS customer success team used this structure for churning customers -- adding a "churn risk signals from the message" field that listed specific language patterns from the ticket that indicated active cancellation intent. Agents responded faster and with better retention language because the draft already addressed the real concern rather than the surface complaint.

A direct-to-consumer e-commerce brand with high seasonal volume used a simplified version of this prompt to build a pre-approved response library for their top 12 ticket types, reviewed by the CS lead once per quarter. The library reduced first-response time by roughly 40% during peak periods because agents weren't drafting from scratch.

An insurance customer service team used the "flag if policy doesn't cover this" instruction specifically to catch edge cases that were previously being handled inconsistently -- different agents making different calls on the same scenario. The flagged cases went to a supervisor for a documented decision, which then fed back into the prompt as explicit policy language for the next round.

Variations for Different Contexts

De-escalation for emotionally escalated messages:

This customer is clearly upset and escalating. Their message: [paste]

Write a response that: (1) acknowledges their specific frustration without defensiveness, (2) takes concrete ownership of one thing we actually did wrong, (3) states exactly what happens next and when.

Don't use the word "apologize" -- it reads as corporate. Don't promise anything we can't actually deliver by [date].

What this does: Banning "apologize" seems counterintuitive, but it forces more specific and credible language -- "I own that this wasn't handled correctly" lands differently than "I apologize for the inconvenience" in an escalated situation.

Three Real Scenarios Where Response Quality Determined Outcome

A subscription software company used the structured response prompt for a situation where a customer had been incorrectly billed for three months without receiving a renewal notice. The specific fields -- customer history (longtime subscriber), desired outcome (full refund plus retain customer), actual policy (refund within 30 days with documented error) -- produced a response that acknowledged the error directly, named the specific refund timeline, and offered a two-month service extension that turned a near-churned customer into someone who left a positive review. The key was the desired outcome field: the agent knew the company wanted to keep this customer, which changed the response from a policy recitation to a genuine repair.

A hardware e-commerce company used the bulk ticket-type approach for their top eight issue categories -- shipping delays, wrong item received, damaged on arrival, return label requests, installation questions, warranty claims, compatibility questions, and cancellations. Each template was built with the actual policy language and a named desired outcome. Response quality scores improved measurably in the first month, not because agents were less skilled, but because they were starting from a better foundation.

A customer success team at a B2B SaaS company used a variant of the customer history field specifically for churn-risk tickets -- tickets from accounts in the bottom quartile of product engagement, where the company knew the relationship was fragile. The prompt instruction was: "this is a churn-risk account -- the response should address both the immediate issue and include one proactive offer that demonstrates we're invested in their success." The proactive offer didn't resolve every churn situation, but it meaningfully changed the pattern for accounts on the fence.

A Prompt Worth Building for Quality Assurance

Review this customer service response for quality: [paste response]

Evaluate against:
1. Does it address the customer's actual issue, or the assumed issue?
2. Does it make any promise that isn't in our actual policy? Flag any commitment that would need verification before sending.
3. Does it tell the customer exactly what happens next and when?
4. Is there any phrase that sounds scripted or corporate? Name the specific phrase.

Rate overall readiness: [ready to send / needs one revision / needs significant rework] with one-line justification.

What this does: Using Claude to QA its own drafts (or a human agent's drafts) before sending catches the specific failure modes -- vague next steps, policy promises that weren't checked, scripted language -- that accumulate over high-volume support shifts when review time is limited. This closes the loop between generation and quality control.

Save and Reuse This

The most valuable thing a customer service team can do with these prompts is treat them as living documents. The first version you write will be good. The version after three months of refinement based on what still needed heavy editing will be significantly better. Storing the current best version somewhere the whole team can access -- a tool like PromptABCD works well for this since it keeps prompt versions organized -- means every agent benefits from every round of improvement, not just the one who made the change.