Client Knowledge Bases
Published pages, ready to be scraped into the AI agent.
How to Build a Client Knowledge Base
We turn a client's web pages into Markdown files, then use AI to read all of those files and write one condensed Markdown file. That single file is what we use for AI Agent training to build the knowledge base.
Capture
Web pages to Markdown
Use a Chrome extension to save each of the client's web pages as its own Markdown file — home, about, services, products, service areas, FAQ, and so on. Any "webpage to markdown" extension will do. This is the one we use:
Get the Chrome ExtensionCondense
Many files into one
Upload all of those Markdown files to ChatGPT (or Claude) together and run the prompt below. The AI reads every file and merges them into one condensed Markdown knowledge base, stripping navigation, buttons, and duplicated boilerplate along the way.
Save the result as Client_Name_Knowledge_Base.md
Publish
Add it to the library
Send the finished Markdown file over and it gets published here as a plain, stripped-down page with no header, footer, images, or styling — just headings and text, so it can be scraped cleanly into the AI agent without picking up junk.
One merged file per client, one client at a time.
The Prompt
Replace [COMPANYNAME] and [WEBSITE] before you send it.
You are building an AI receptionist knowledge base for [COMPANYNAME] ([WEBSITE]). I am attaching several Markdown files. Each one was converted from a single page of this company's website. Read all of them and merge them into ONE condensed Markdown file that will be used to train our AI agent. Rules: - Use only facts found in the source files. Never invent, infer, or estimate anything — especially pricing, warranties, response times, or availability. - Strip all navigation text, buttons, cookie notices, image captions, form labels, and repeated boilerplate. - Deduplicate. If the same fact appears on several pages, state it once. - Attribute soft or promotional claims to the source: "The website states that..." rather than asserting them as fact. - Where the source is incomplete, write an instruction for the agent instead of guessing. For example: "Details should be confirmed with the company." - Keep phone numbers, addresses, and the website domain exactly as written. - Use plain sentences an agent can say out loud. No marketing language, no emojis, no tables, no images, no links. Use these headings, skipping any that the sources do not support: # [COMPANYNAME] — AI Receptionist Knowledge Base ## Company Overview ## Services ## Products & Capabilities ## Service Area ## Common Questions ### (one H3 per question, with the answer in plain text beneath it) ## Consultation & Quotes ## Company Background & Proof Return raw Markdown only, with nothing before or after it.