Build a help centre out of the answers you have already sent

Turn two years of support replies into a searchable help centre, reviewed by you before anyone sees it.

For anyone answering the same question in a dozen threads · 8 steps · 7 min

How it works today.

You have answered the same five questions a hundred times. The wording shifts slightly each time—different customer, different mood, different product version—but the core answer is stable. Someone asks about billing cycles, and you search your sent folder for the last time you explained it, then retype half of it because the old reply mentioned a feature you retired.

You know a help centre would cut this volume in half. But writing articles from scratch means blocking out days you do not have, and by the time you finish the first draft the product has moved on. So you keep replying one by one, and the good explanations stay buried in threads only you can see.

Before you start.

All of it has to be true, or step one fails in a way that is annoying to debug.

Export of your support replies from the past 12–24 months, in a format that preserves question and answer pairs (CSV, JSON, or connected via API).
Access to a generalist agent that can read documents and generate structured text (ChatGPT, Claude, or equivalent).
A staging or draft area in your help centre tool where articles can sit unpublished until you approve them.
Write access to publish in your help centre, or a clear handoff to whoever does.

The steps.

Export your support replies with enough context to reconstruct the question
Pull tickets or threads from the past year where you marked the conversation resolved. Include the customer's original message and your reply, plus any follow-up if it clarified the answer. If your tool lets you filter by tag or macro, start with your most-used categories. Save as CSV or connect the agent directly if your platform has an API the agent can read.
02
Group replies by the question they actually answer
Feed the export to the agent and ask it to cluster replies by intent, ignoring surface wording. You want to see which answers are solving the same underlying problem, even if one customer asked about invoices and another asked about receipts. The agent should return a list of question themes with sample replies under each.
Paste this
```
Read these support replies and group them by the core question the customer was trying to answer. Ignore politeness and phrasing differences. Return a list of themes, each with 3–5 example replies that belong to it. Show me which questions we answer most often.
```
Pick the five themes that would cut your reply volume the most
Look at the grouped output and choose the questions you answer weekly or daily. Ignore one-off edge cases and anything tied to a feature you are about to retire. You are looking for stable, high-frequency answers that will stay true for at least six months.
04
Draft one help article per theme, synthesising the best parts of your past replies
For each theme, ask the agent to write a help centre article that combines the clearest explanations from your replies. The agent should use your actual words where they work and smooth out contradictions where the answer evolved. Specify your help centre's tone and structure if you have a style guide. The output should be draft HTML or Markdown, ready to paste into your CMS.
Paste this
```
Write a help centre article for [theme title]. Use these replies as source material: [paste the 3–5 examples]. Keep the tone [your brand voice notes]. Structure: short intro, step-by-step answer, one-sentence summary at the end. Output as Markdown. Do not add information that is not in the replies I gave you.
```
Check every claim in the draft against your current product
Open each draft and read it next to your live product or internal docs. Confirm that field names, screen locations, and behaviour still match. If the agent synthesised two replies from different product versions, you will catch the stale one here. Mark anything wrong and either fix it yourself or send it back to the agent with the correct detail.
Paste the draft into your help centre as an unpublished article
Copy the Markdown or HTML into your CMS and save it as a draft. Add screenshots if your help centre expects them—this is still your job. Set the article slug and category so it will appear in the right section when you publish. Do not publish yet.
Send the draft link to one person who asks this question often, and watch where they get stuck
Pick a customer or teammate who has asked this question in the past month. Send them the draft link and ask them to try following it. Note where they pause or ask a follow-up question. If they succeed without coming back to you, the article works. If they get confused, revise that section before publishing.
Publish the article and link to it the next time someone asks
Make the article live in your help centre. When the question comes in again, reply with a one-sentence answer and the link. If the customer replies saying the article did not cover their case, note what was missing and add it in your next review cycle. Track how often you send the link versus writing a new reply—that ratio tells you if the article is working.

What you keep.

Automating the typing does not move the accountability. These stay with a person.

Deciding which answers are stable enough to publish. The agent groups replies, but you know which questions have settled and which are still shifting with product changes.
Reading every draft article against the live product to confirm it is still true. The agent synthesises your past replies, but it cannot check whether a field was renamed last sprint.
Approving publication. A help article is a promise to every customer who reads it, and you are the one who knows if that promise holds.

Once it works.

The first run is the demo. These are where the time actually comes back.

Run it on a loop

Run this process quarterly on your newest resolved tickets to catch questions that have become frequent since the last build.

Fire it from an event

When you answer the same question three times in one week, flag those replies for the agent to draft an article immediately instead of waiting for the next batch.

Scale it wider

Once you have fifteen published articles, ask the agent to generate a single-page FAQ by pulling the one-sentence summaries from each, so customers can scan the full list before searching.

Hand off to another agent

If your help centre has a feedback widget on each article, route low-rated pages back to the agent with the feedback text and ask it to suggest a revision for you to review.