Build Your Support Agent
Create a SUPPORT.md that tells an agent how to handle customer questions, then wire it up to draft support replies, monitor for churn signals, and surface issues that need your attention.
Why This Matters
Support is the moment of truth. A customer who gets a fast, accurate, human-sounding reply is more likely to stick around than one who waits 48 hours for a canned response.
Most early-stage founders handle support personally — and should. You learn from every support conversation. But you also spend 2-3 hours a day on questions that have the same 10 answers.
A support agent doesn't replace you. It drafts replies to the 80% of tickets that have clear answers, flags the 20% that need your judgment, and monitors for the signals that precede churn. You review, approve, and send. Response time drops to minutes. You stay in the loop on what matters.
What You're Building
SUPPORT.md — The support brain:
- How to handle the 10 most common questions (with approved answers)
- Escalation rules (what gets human review immediately)
- Refund and cancellation policy (exact language)
- Tone and empathy guidelines
- What to never say (no promises about features, no blame, no corporate-speak)
- Churn signals to flag (multiple failed payments, login drops, downgrade requests)
A working support draft workflow — that reads tickets, drafts replies, and surfaces escalations.
Implementation Tiers
Tier 1: Claude Projects (15 minutes)
Add SUPPORT.md to your Claude Project. Use it as context for a dedicated support conversation where you paste tickets and get draft replies.
Step 1: Generate SUPPORT.md
Based on COMPANY.md, write SUPPORT.md — the support playbook for an AI agent handling customer questions for [product name].
Include:
1. The 10 most common support questions for a [product category] SaaS — with our approved answer for each
2. Escalation rules — which tickets get immediate human response (billing disputes, data questions, threats to churn, angry tone)
3. Refund policy — exact language we use, under what conditions we offer refunds, what we say when we can't
4. Tone guide — how to open replies, how to close them, how to be empathetic without being sycophantic
5. What we never say — no promises about roadmap, no admitting fault for service outages without authorization, no competing-product comparisons
6. Churn signal definitions — patterns that indicate a customer is about to cancel
My product's common support scenarios:
- [scenario 1, e.g., "Can't connect my account"]
- [scenario 2, e.g., "Why was I charged X"]
- [scenario 3, e.g., "How do I export my data"]
Step 2: Draft support replies
Here is a support ticket:
Subject: [subject]
From: [customer name]
Plan: [their plan]
[paste ticket body]
Using SUPPORT.md and COMPANY.md:
1. Classify this ticket: routine / needs-escalation / churn-risk
2. Draft a reply that follows our tone guidelines exactly
3. If escalation is needed, explain why and what I should do
4. Flag any SUPPORT.md gaps — if you couldn't find an approved answer, say so
Step 3: Build your reply library
As you review and send replies, note which ones work best. These become approved templates in SUPPORT.md.
Tier 2: Claude Code (30 minutes)
Wire Claude Code to your support inbox via a script that pulls tickets, drafts replies, and outputs a review queue.
Step 1: Create SUPPORT.md
# context/SUPPORT.md generated using prompt above
Step 2: Simulate the support queue
I have these support tickets that came in today:
1. [paste ticket 1]
2. [paste ticket 2]
3. [paste ticket 3]
For each ticket, using SUPPORT.md and COMPANY.md:
1. Classify: routine / escalation-needed / churn-risk
2. Draft a reply (ready to send, no editing needed if routine)
3. Note confidence level: high / medium / needs-human-review
Output to support/queue-[date].md as a structured table with the draft reply in each row.
Step 3: Build a metrics alert script
Write a script that reads support/queue-[date].md and outputs:
- Ticket count by classification
- Average estimated resolution time
- Any patterns in churn-risk tickets (common complaints)
Save to support/weekly-summary.md
Tier 3: Paperclip (1 hour)
A Paperclip support agent connects to your support inbox, classifies tickets, drafts replies, and surfaces escalations — with routine replies going out after a 2-hour human review window.
Step 1: Wire SUPPORT.md to the agent
Add to your Support Agent's AGENTS.md:
## Support Agent Instructions
Before any support task, read:
- `COMPANY.md` — product, pricing, policies
- `SUPPORT.md` — approved answers, escalation rules, tone, refund policy
For each support ticket:
1. Classify: routine / escalation-needed / churn-risk
2. Draft a reply following SUPPORT.md tone guidelines exactly
3. For routine tickets: create a task in the review queue with the draft
4. For escalation/churn-risk: immediately create a high-priority task and @-mention the board
5. Track ticket patterns — if the same question appears 3+ times this week, flag it as a potential FAQ gap
Hard rules:
- Never send a reply directly — always create a review task
- Never promise features or timelines not in COMPANY.md
- Never offer a refund without board approval unless within stated policy
Step 2: Set up the review queue
Create a Paperclip project for Support. All drafted replies appear as tasks in this project. You review them in the morning and evening — 2-4 minute sessions instead of an hour of inbox work.
Step 3: Monitor churn signals
Configure the agent to run a weekly report:
Every Sunday at 6pm: review all tickets from the past 7 days.
Identify: churn signals, recurring complaints, feature requests.
Create a weekly summary task assigned to the board.
You get a structured weekly brief on what's bothering your customers — without touching individual tickets.
Deliverable
SUPPORT.md— complete support playbook with approved answers and hard limits- A support draft workflow that reduces your ticket time by 70%
- A churn signal monitoring process
See the SUPPORT.md template for a starting point.
What's Next
With your support and conversion systems running, move to Agent Company Goes Live — where you launch publicly, hand off operations to your agent team, and establish the rhythms of a sustained one-person company.