A ready-to-use system prompt for an internal knowledge base agent that answers your team's questions from your wiki, runbooks, and docs, cites the source, and says when something isn't documented. Copy it, fill the brackets, and paste it into your agent's instructions.
What this agent does
Internal knowledge is scattered across a wiki, runbooks, onboarding docs, and old threads, and people waste time asking colleagues what's already written down. This prompt makes the agent answer team questions from that indexed content, citing the exact page so the answer is verifiable and the asker can read more. It's strict about grounding — when the knowledge base doesn't cover something, it says so, surfacing the gap instead of inventing an answer that could mislead the team.
The prompt
You are the internal knowledge assistant for [Team]. Answer questions
only from the provided context drawn from [Team]'s wiki, runbooks, and
docs, and cite the source page. If the answer is not in the context, say
so clearly so the gap can be filled. Be concise and direct; assume an
internal audience. Never invent processes, owners, or policy.
How to use it
Replace [Team] with your team or company name, index your wiki and docs (a crawl, uploaded files, or a connected source), and paste the prompt into your agent's instructions. The analytics dashboard shows the questions your knowledge base doesn't answer — a backlog for your docs owners. See the SaaS use case for the broader pattern, then start free.
Why grounding beats a bare prompt
A prompt alone doesn't make a knowledge base agent reliable — grounding does. Pasted into a generic chatbot, these instructions still let the model answer from its general knowledge and invent details. In AIML.chat the same prompt sits on a retrieval layer that feeds the agent only the content you indexed, so "answer only from the provided context" is enforced by the system rather than merely requested. Every claim links back to a page you control, and the analytics record what people ask so you can refine both the prompt and the content behind it. That pairing — your instructions plus your own grounded content — is what turns this template into a dependable agent.
What can it index?+
A crawl of your wiki, uploaded files, a GitHub repo, or a connected MCP source — whatever holds your internal knowledge.
Will it invent a process when it doesn't know?+
No — the prompt instructs it to say the answer isn't documented, surfacing the gap rather than guessing.
Is there a free plan?+
Yes — 100 messages a month and one website on the free tier. See pricing for higher limits.