Custom AI Assistant · Lesson 2Pro~14 min readDocs + voice + guardrailsBuilds on Lesson 1

A real company brain — accurate, on-voice, in its lane.

A one-doc assistant is easy; a knowledge hub your team trusts is a craft. This lesson is about curation and guardrails — loading real documents, giving the assistant your voice and a never-say list, and making every answer trustworthy with citations and honest gaps.

The mental model

Scaling from one job to a company brain is about curation and guardrails — what you feed it, and the rules you give it.

A one-doc assistant is easy. A team’s knowledge hub is only as good as the documents behind it and the rules that keep it honest. The work shifts from writing instructions to curating knowledge and setting guardrails so it stays accurate, on-voice, and in its lane.

The Reframe

Garbage in, confident garbage out. The assistant will faithfully reflect whatever you feed it — including outdated or contradictory docs. Curation is the job now.

Step 01 Load real documents

Move from one file to a proper knowledge base: handbooks, product docs, FAQs, past answers. Structure helps — clear titles and current versions so the assistant can find the right thing.

Step 02 Give it voice and guardrails

A company brain represents you, so set the rules explicitly:

The guardrail set

Voice guide — how your company sounds (from your best existing writing).
Do / don’t — what it should and shouldn’t do.
Never-say list — topics or claims it must avoid (pricing promises, legal/medical advice, competitors).
Escalation rules — exactly when to hand off to a person.

Step 03 Make answers trustworthy

Trust comes from honesty, not confidence. Instruct it to:

Cite the source — name the document an answer came from.
Admit gaps — say when something isn’t covered.
Stay in scope — decline cleanly when a question is outside its job.

Trust promptWhen you answer, cite which document it came from. If the documents don’t cover the question, say so and don’t guess. If it’s outside your scope ([scope]), say it’s not something you can help with and point to [where].

Step 04 Keep it current

Knowledge rots. Decide who owns updates and how often, and remove or replace outdated docs — a confidently wrong answer from a stale policy is worse than no answer.

Two traps: conflicting documents (the assistant picks one at random — keep a single source of truth per topic), and oversharing (it’ll happily surface internal or sensitive info to anyone who can use it — control what goes in, and see the Secure Your AI Use build).

Your challenge: build a trustworthy brain

Grow your Lesson 1 assistant into a real one:

Load several real documents as a structured knowledge base.
Add a voice guide, a do/don’t, and a never-say list.
Make it cite sources and admit gaps; test the citations.
Decide who owns updates and how often.

That’s an assistant your team can actually trust. Next, deploy it with proper access controls, connect it to live data, and keep it current — that’s Lesson 3.

Next: share it with your team →Back to Custom AI Assistant

What you can do now

Build a structured, multi-document knowledge base
Give an assistant a voice guide and guardrails
Write a never-say list and escalation rules
Make answers cite sources and admit gaps
Keep knowledge current and avoid conflicting docs

Pro+

Up next in Custom AI Assistant

Lesson 3 · Share it with your team — access, live data & upkeep

Deploy it with the right access controls, connect it to live data instead of static files, and keep it current. Go to Lesson 3 →