A Markdown knowledge base you can reuse across AI assistants

September 5, 2026 · Ask Your Docs

A Markdown knowledge base keeps reference material in readable text files with headings, links and explicit source ownership. You can move that content between tools, but file portability does not guarantee synchronized updates, correct AI answers or equivalent access controls. Build a small collection, test its rules and exceptions, then expand around actual work questions.

What belongs in a Markdown knowledge base?

Markdown works well for policies, procedures, product definitions, small comparison tables and explanatory notes. A maintained reference might answer who approves an offer, which process applies to a request or where to find the authoritative source. Headings help separate scope, rules, exceptions and related material.

The format is less suitable as the only record of a financial model, a signed agreement or a presentation whose meaning depends on diagrams. Keep those originals and write a short reference explaining their purpose. A text export cannot preserve a spreadsheet’s calculation engine or a document’s complete review history.

For an existing Word reference, the DOCX converter can provide a starting point. Use the Excel and CSV converter for a selected table. Compare the exported text with the source, particularly where an exception appears outside a table. Ask Your Docs also accepts original file formats; Markdown is an organizational choice, not a required upload step.

Start with a small, connected example

The download describes a fictional Cedar sales team. It is separate from the Harbor campaign examples elsewhere on this site. No policy, company or named owner in the exercise represents a real customer. The corpus is intentionally small enough to inspect before you ask an assistant to use it.

The pack also includes a README and evaluation CSV. Upload only the four source files into the collection being tested. Keep the evaluation answer sheet outside it. Otherwise the assistant could repeat a prepared answer without retrieving the relevant policy.

Give every source an owner, version and scope

Each example file identifies an owner, version, review date and status. These are human-readable fields near the beginning of the document. They make it easier to compare an answer with the intended source version without requiring a particular metadata parser.

For your own library, define what the source covers and what it does not. The discount policy explicitly excludes refunds, renewals, tax and currency conversion. This boundary supports a useful missing-information answer. It also prevents a rule for new annual subscriptions from being presented as a universal commercial policy.

Metadata is descriptive. Writing “restricted” in a Markdown file does not prevent someone from reading it. Apply permissions in the storage and assistant systems, and check access with the intended audience. Similarly, a review date is not proof that a review took place; an owner must perform and record that work.

Use links without assuming automatic retrieval

The files contain ordinary relative links. When the folder is kept together in a compatible editor, those links connect the policies and directory. If you upload a single file into another tool, the linked file may not be available. An index is a map, not a substitute for the source content.

Upload or connect all four sources explicitly. Then ask a question that requires both a policy rule and a role lookup. A correct role without the right person, or a person without the governing approval rule, signals that the answer did not combine the required sources completely.

Choose filenames that describe their contents and remain stable across routine edits. Record version changes inside the source and in your source register. If a filename must change, update referring links and confirm that the imported collection no longer depends on the old path.

Run the same questions in ChatGPT and Claude

ChatGPT Projects supports project reference material and instructions. Anthropic documents adding project knowledge and instructions in its Claude project setup guide. Use the current controls available to your account, and confirm all four files are present before testing.

Add a short instruction asking for answers from the supplied sources, with a filename, heading and version. Ask the assistant to separate requests from approvals, apply exceptions and admit missing facts. The ChatGPT guide and Claude guide provide more setup detail.

1. Who must approve a 12% discount on a new annual subscription?
2. Can I put a customer quotation in a deck after discount approval?
3. How long is a commercial quote valid?
4. What is the refund policy?
5. Does naming Morgan Lee in a request prove finance approved it?

Record the actual answer and source beside the expected answer in the evaluation CSV. These are authored checks, not reported ChatGPT or Claude test results. Keep the assistant, model, account context, date and source version in a run log so another person can understand the comparison.

Inspect evidence separately from wording. A fluent answer that names both approvers but cites only the directory has not shown the threshold rule. An answer that treats a commercial quote as an attributed customer quotation has confused two distinct concepts. Source checks should catch these errors before the corpus becomes operational guidance.

Test a source update, not just the first upload

In a copy of discount-policy.md, change commercial quote validity from 14 to 7 calendar days and increase the version from 1.0 to 1.1. Replace the active reference in the assistant’s collection. Keep historical copies out of the current collection unless they are clearly identified as historical and needed for a specific question.

Start a fresh conversation and ask the validity question again. The updated expected answer is 7 days, supported by version 1.1. If an answer still says 14 days, inspect the actual connected files and returned passage. Do not assume the assistant is using the local file you just edited.

Repeat the update separately in the second assistant or verify the shared-library connection that supplies both. A Markdown format does not provide synchronization by itself. A reliable maintenance process must specify who updates the source, how the active copy changes and which questions are rerun.

Grow around questions your team actually asks

Add a reference when a recurring question has no maintained answer. Keep one authoritative source for a rule rather than copying it into every related document. Link to it and explain the relationship. Duplicate rules create a maintenance problem when one copy changes and another does not.

Use the knowledge base starter kit for a source register and update log. The broader AI knowledge base guide explains how a mixed-file library fits the four-step workflow: collect files, organize sources, connect your AI and ask with evidence.

Frequently asked questions