Build a Company Q&A Assistant with a Custom GPT

What You'll Learn

• The exact instruction prompt to configure a company knowledge GPT
• What documents to upload and how to structure them for best GPT performance
• How to test the GPT before sharing it with your team to ensure accuracy

A head of HR at a 150-person technology company answers the same seven questions every week. What is the expense reimbursement limit for client meals? How many days notice are required for PTO? Where is the performance review template? She knows the answers. The answers are in the handbook. The handbook is a 60-page PDF that nobody reads.

A custom GPT with the handbook uploaded answers all seven questions in under 5 seconds, accurately, any time of day. Building it takes 20 minutes. Below is the exact prompt to configure it.

SUBSCRIBER BREAK -- Premium Content Below

What to Prepare Before Starting

Source documents: Employee handbook, expense policy, PTO policy, onboarding guide, existing FAQs. Maximum 512MB per file, up to 20 files per GPT. Well-structured documents with clear headers produce significantly better outputs than unformatted text blocks.

The 10 most common questions: Write them down. These become your test cases after the GPT is built.

Scope boundaries: Decide what the GPT should NOT answer. Policy exceptions, legal advice, and personnel decisions should always route to a human. Define these limits in the instructions.

The GPT Instruction Prompt

Paste this into the "Instructions" field in the GPT Builder Configure tab. Customize the bracketed sections for your organization.

You are [COMPANY NAME]'s internal knowledge assistant. Your role is to help employees and managers quickly find accurate answers to questions about company policies, procedures, and guidelines. KNOWLEDGE BASE: You have access to [list uploaded documents, e.g., the Employee Handbook, Expense Policy, PTO Guidelines, and Onboarding Checklist]. Answer all questions using these documents as your primary source. HOW TO ANSWER: - Give a direct, plain-language answer first - Cite the specific section or page of the document where the answer can be found - Keep answers under 150 words unless a longer explanation is genuinely necessary - Use bullet points for multi-part answers WHEN UNSURE: - If the answer is not clearly in your documents, say: "I don't see a clear answer to this in the current documentation. For this type of question, I recommend contacting [HR / Legal / Finance] directly." - Do not guess or extrapolate beyond what the documents state - Do not provide advice on personnel decisions, legal interpretations, or policy exceptions SCOPE LIMITS: You are a reference tool, not a decision-maker. For questions involving exceptions to policy, disciplinary matters, benefits enrollment, or legal questions, always refer the user to the appropriate human contact. TONE: Clear, friendly, and professional. Write as if explaining to a colleague who is in a hurry.

Building and Testing the GPT

• Go to chat.openai.com, click your profile picture, select "My GPTs," and click "Create a GPT." Click "Configure."
• Paste the instruction prompt above (customized) into the "Instructions" field.
• Under "Knowledge," upload your source documents. Wait for each file to finish uploading before adding the next.
• Set "Conversation starters" to your three most common questions. This helps new users understand what to ask immediately.
• Click "Preview." Test all 10 of your most common questions. Verify: Is the answer accurate? Does it cite the source? Does it route out-of-scope questions to a human?
• For any test that fails, return to Instructions and add a specific clarification. Save and re-test.
• Set sharing to "Only people with a link." Share with your team with a one-paragraph explanation of what the GPT is for and what it should not be used for.

Why This Prompt Works

Three things most generic GPT setups omit: (1) explicit citation requirement, so users can verify answers; (2) a defined fallback when uncertain, routing to a human instead of hallucinating; (3) explicit scope limits, so the GPT stays accurate within its domain. The 20-minute investment typically returns 2-4 hours per week in reduced interruptions for the team members who currently field these questions.