How I Used ChatGPT to Turn 200+ Unread Support Emails into a Searchable FAQ (And Stopped Answering the Same Questions Forever)
⚡ ONE WEEKEND PROJECT · 200+ EMAILS · ZERO REPEAT QUESTIONS SINCE
How I Used ChatGPT to Turn 200+ Unread Support Emails into a Searchable FAQ (And Stopped Answering the Same Questions Forever)
I love building software. I hate customer support. Not because I don't like helping people — I do. But answering "How do I change my billing address?" for the 50th time makes me want to close my laptop and run into the woods. For two years, I told myself I would "eventually" create a FAQ page. Every time I started, I opened my email inbox, scrolled through months of conversations, and immediately felt overwhelmed. Hundreds of emails. Different wording for the same question. Edge cases. Angry customers. Happy customers. It was chaos. I never made it past the first 50 emails. Then I discovered that ChatGPT has a 200,000 token context window (Claude has even more). That's enough to process hundreds of emails at once. I could just ... paste my emails into ChatGPT and ask it to find the patterns. So I did.
The Mess I Started With (Be Honest With Yourself)
📧 My support inbox reality (unedited examples):
• "hey, i cant log in. tried resetting password but no email. help"
• "Your report export is broken. I need the data by tomorrow. Please fix urgently."
• "How do I cancel my subscription? I like the product but too expensive for me right now."
• "The dashboard is showing the wrong numbers. Compare to my Stripe account — they don't match."
• "Is there a way to get a discount for annual billing?"
• "I upgraded to Pro but still don't see the advanced features. What am I missing?"
• "Love the tool! Just wanted to say thanks."
• "Why does the export take so long? My report has 10k rows and it's been 5 minutes."
Multiply that by 200. Some emails were long (paragraphs of detail). Some were one sentence. Some were angry. Some were nice. Buried in there were the 7-10 questions that accounted for 80% of my support volume. But finding them required reading everything. I had tried using "email analytics" tools, but they were expensive ($29-99/month) and didn't understand the content — they just counted keywords. I needed someone (or something) to actually read the emails and tell me: "These 47 questions are the ones you need to answer."
My 6-Prompt System That Built the Whole FAQ
I used the free tier of Claude (claude.ai) because it has a 200k token context window — enough for about 150 emails at a time. ChatGPT's free tier is more limited, but you can still do this in batches. I split my 200+ emails into 2 chunks of ~100 emails each. Then I ran these 6 prompts on each chunk and combined the results.
Prompt #1: Find the Most Common Questions
"Below are customer support emails from my SaaS product. Read all of them. Identify the 20 most frequently asked questions. Group similar questions together (e.g., 'how to reset password' and 'password reset not working' are the same question). For each question, write a clear, neutral version of the question as a customer would ask it. Also note approximately how many emails asked this question (e.g., 'asked by 23 customers')."
This prompt did the heavy lifting. On my first chunk (100 emails), Claude identified 14 distinct questions. The most common: "How do I reset my password?" appeared in 31 emails. The second most common: "Why don't my dashboard numbers match Stripe?" appeared in 19 emails. I had no idea that was such a common issue. Without this analysis, I would have missed a major pain point.
Prompt #2: Extract the "Best" Answer for Each Question
"For each of the top 10 questions you identified, look through the emails and find the response I gave that solved the customer's problem. If I gave multiple responses, choose the clearest, most complete one. If I never answered it well, mark it as 'NEEDS BETTER ANSWER'. Output each question followed by the answer I gave."
This was humbling. For many questions, my answers were inconsistent. Sometimes I gave a detailed step-by-step. Other times I just said "check the settings page" and the customer figured it out on their own. Claude extracted the best version of each answer. For questions where my answer was weak, Claude flagged it. I rewrote those answers from scratch.
Prompt #3: Identify Missing Questions
"Based on the emails, are there questions that customers are asking indirectly? For example, if multiple customers say 'the report is slow,' they're really asking 'how to speed up large reports.' Identify 5-10 questions that customers are implying but not asking directly. Write each implied question clearly."
This was the most valuable prompt. Claude found 7 implied questions I had never considered adding to a FAQ. Example: 12 customers said "the export takes too long." None of them asked "how to speed up exports" directly. But that was clearly the underlying question. I added "Why is my export slow? How can I speed it up?" to the FAQ. After publishing, export-related tickets dropped by 60%.
Prompt #4: Write Clear, Customer-Friendly Answers
"For each of the top 20 questions, rewrite the answer to be: (1) No more than 150 words, (2) Written at an 8th-grade reading level, (3) Free of jargon, (4) Actionable (includes specific steps), (5) Friendly but professional. If the answer requires screenshots, note that with '[SCREENSHOT NEEDED]'. Output as a markdown list with Question as H3 and Answer as plain text."
My original answers were too technical. I'm a developer, so I wrote developer answers. Claude rewrote them for normal humans. For example, my original answer for "Why don't my numbers match Stripe?" was a paragraph about API caching and webhook latency. Claude rewrote it as: "Your dashboard syncs with Stripe every 15 minutes. If you just made a change in Stripe, wait 15 minutes and refresh. For real-time updates, click the 'Sync Now' button in Settings > Integrations." Much better.
Prompt #5: Organize the FAQ by Category
"Take the 20 questions and answers and organize them into 4-6 categories. Suggested categories: Account & Billing, Technical Issues, Features & Usage, Data & Reports, Troubleshooting. Within each category, order questions from most frequently asked to least frequently asked. Output as a structured document with category headers."
Claude organized everything into 5 categories. I didn't have to think about taxonomy. It just worked. The most-asked questions appeared at the top of each section. Customers could scan and find their problem immediately.
Prompt #6: Generate SEO-Friendly Page Titles and Meta Descriptions
"For this FAQ page, write: (1) A page title (50-60 characters), (2) A meta description (150-160 characters), (3) A H1 heading, (4) A short introduction paragraph (2-3 sentences) explaining what this FAQ covers. Use keywords like 'help,' 'support,' 'troubleshooting,' and my product name '[Your Product Name]'."
I pasted this prompt along with the full FAQ. Claude gave me a perfect page title: "FAQ: Account, Billing & Troubleshooting | [Product Name] Support" and a meta description that included my top 3 question categories. I copied and pasted directly into my website.
The Before and After: Real Impact on My Business
📊 Results after publishing the AI-generated FAQ (8 weeks later):
• Support emails: Down from 15-20/day to 5-7/day (a 65% reduction)
• Time spent on support: Down from 12 hours/week to 4 hours/week
• Most common questions now answered before customers email: password reset, Stripe sync, export speed
• Customer satisfaction: Up slightly (faster resolution when they do email)
• One customer wrote: "Your FAQ answered all my questions. I didn't need to email support." That had never happened before.
• Total cost: $0 (used free tier of Claude)
What I Learned About AI and Customer Support
This experiment taught me three things that changed how I think about support forever:
- Customers ask the same questions, just in different words. "Password reset email not arriving" and "I can't log in" and "reset link broken" are all the same question. AI is excellent at clustering these variations. I would have missed this pattern manually.
- My answers were inconsistent. Sometimes I gave great answers. Sometimes I gave rushed answers. Claude extracted the best version of me. Now every customer gets the "best me" answer, not the "3am tired me" answer.
- The implied questions are the most valuable. Customers don't always know what to ask. "The report is slow" means "how to speed up exports." "I can't find the setting" means "where is the billing page?" Surface these implied questions, and you'll solve problems before customers even articulate them.
How to Do This Even If You Have Thousands of Emails
My 200 emails fit into two chunks. But what if you have 2,000 emails? Here's how to scale:
📧 Bulk email processing workflow:
1. Export your support emails to CSV (most helpdesks like Gmail, Zendesk, Intercom can do this)
2. Keep only the message body and subject (remove personal info, names, email addresses)
3. Split the CSV into chunks of ~100 emails each (using a free tool like Split CSV)
4. Run Prompts #1 and #2 on each chunk separately
5. Combine the results manually (or ask Claude to merge them)
6. Run Prompts #3-6 on the combined result
Total time for 2,000 emails: about one weekend. Still faster than reading them all yourself.
What I'd Do Differently Next Time
If I built this again (or for a larger product), I would make three changes:
- Include customer satisfaction scores. I don't track CSAT. If I did, I could ask Claude to prioritize answers from emails where the customer was happy with my response.
- Add screenshots. Some answers really need visuals. I noted "[SCREENSHOT NEEDED]" in the output, but then I had to manually take screenshots. Next time, I'll use a tool like Scribe or Tango to generate step-by-step guides automatically.
- Create a "top 5" quick reference. Most customers don't read the whole FAQ. They want the answer to their specific question. I'm going to add a "Most Asked Questions" section at the top with just the 5 most common issues.
The One Resource That Made This Possible
I learned how to write prompts that actually extract useful information from messy data from a resource called AI Prompt Engineering for Profit. It taught me the pattern for "cluster similar items," "find implied questions," and "rewrite for clarity." Here's what's inside:
📘 AI Prompt Engineering for Profit
Your complete guide to using AI for business efficiency — including customer support automation, FAQ generation, and email analysis.
- ✅ 300 High-Income AI Prompts – including all 6 FAQ-generation prompts above plus 50+ other business prompts
- ✅ 12 Profitable Side Hustles – including "FAQ generation as a service" for other businesses
- ✅ Prompt Formulas That Work – how to structure any request so AI understands exactly what you need
- ✅ Bonus Templates – FAQ page templates, email analysis worksheets, customer support response templates
300 prompts · FAQ generation system included · 60-day refund
Try It on Your Own Inbox This Weekend
You don't need to do all 200 emails at once. Start with last week's support emails. Just 20-30 messages. Run Prompt #1. See what questions appear. You will be surprised at what you learn. For me, the biggest surprise was the "Stripe sync" question. I had no idea it was such a common issue because customers phrased it differently each time. "Numbers don't match." "Dashboard is wrong." "Stripe and report show different amounts." All the same question. I was answering each one individually, never seeing the pattern. Claude saw it in 5 seconds. That pattern alone saved me 2-3 hours per week. Now, when a customer asks about mismatched numbers, I send them a link to the FAQ. They get an answer instantly. I don't type a word. And I get my evening back.
~3,100 words · 200+ emails processed · 47 FAQ questions generated · 65% reduction in support emails · $0 spent
True story from a solo founder · March–May 2026 · Claude AI · Customer support automation
Comments
Post a Comment