Build a Gem

A Gem is a custom version of Gemini that follows your instructions and answers from files you point it at. You can build one in an afternoon without asking anyone’s permission.

Good for

  • Answering your team’s recurring internal questions
  • Onboarding new staff
  • Policy and procedure questions for employees
  • Testing whether a bigger chatbot idea is worth pursuing

Not good for

  • Anyone without an ND Google account. If your audience includes prospective students, families, or the public, a Gem won’t reach them.
  • Living on your public website
  • Telling you much about how it’s being used

Before you start

  • Know exactly who it’s for
  • Gather the documents it should answer from
  • Check the data rules below
  • Write down ten real questions people actually ask

How to build it

  1. Sort out your content first

    This is where the quality is won or lost, and it’s the step everyone wants to skip. Retire anything out of date, resolve documents that contradict each other, and make sure each one makes sense on its own. How to prepare content.

  2. Point it at your files in Drive

    Link to the documents where they already live rather than uploading separate copies. When someone updates the source document, the Gem works from the updated version — you don’t have to remember to re-upload anything. Check the sharing on those files while you’re there.

  3. Write the instructions

    Say who it’s for, what it covers, what it should refuse to answer, roughly how long answers should be, and what tone to take. Be specific about what’s out of scope — that’s the instruction people most often leave out. See below for a faster way to draft this.

  4. Test it

    Start with your ten questions. Then try ten you didn’t write — ask a colleague to have a go, because they’ll phrase things in ways you wouldn’t.

  5. When an answer is wrong, look at the content before the instructions

    Most wrong answers come from missing, stale, or contradictory source material. Rewriting the instructions won’t fix that.

  6. Share it deliberately

    Share with named people or a specific group. Don’t share broadly by default, and remember that people need access to the underlying Drive files too.

Let an AI draft the instructions

Writing instructions from a blank box is harder than it needs to be. Describe what you want to an AI assistant and have it produce a first draft, then edit that.

Something like this works well as a starting prompt:

I’m building a chatbot for [who it’s for] at the University of Notre Dame. It should answer questions about [topics] using documents I’ll provide. It should not attempt [what’s out of scope]. When it doesn’t know an answer it should say so and point people to [contact or page]. Answers should be [length and tone]. Draft a set of instructions for it. Ask me any questions you need answered first.

That last sentence does most of the work. An assistant that asks you about your escalation contacts, your sensitive topics, and your edge cases produces far better instructions than one that guesses.

Two things to watch:

  • Edit what comes back — don’t paste it unread. The draft won’t know your deadlines, which topics are sensitive, or who questions should be routed to.
  • Bring your test failures back to it. “When I asked X it answered Y, which is wrong because Z — how should the instructions change?” is a fast way to improve them, as long as you check the answer isn’t really a content problem.

The same approach works for your greeting message and your starter questions.

The data rules still apply

No one has to approve your Gem. That is not the same as there being no rules.

Use your unit’s own content freely. For data another unit owns, or anything restricted — student records, health, financial, or personnel information — get that data’s steward to approve the use first.

Don’t share the Gem, or the files behind it, more widely than the content allows. A document that’s fine for your team isn’t automatically fine for the whole University.

This is the same rule that applies to a managed chatbot. Building it yourself doesn’t change it.

Keeping it current

Pointing at Drive files solves the mechanical half of this: update the document and the Gem uses the updated version. It doesn’t solve the other half, which is that somebody has to notice the content needs updating in the first place.

  • Put a recurring reminder in your calendar — quarterly is a reasonable default
  • Add a “current as of [date]” line to your instructions so users know how fresh it is
  • Re-run your test questions after any significant content change
  • Check that the files it points to haven’t been moved, renamed, or had their sharing changed

A Gem that quietly went stale six months ago is worse than no Gem, because people still trust it.

When to come to us instead

Move to a managed chatbot when:

  • You need to reach people without ND accounts
  • People outside your unit are asking to use it
  • Your content set has grown beyond what a Gem handles comfortably
  • Your team has come to depend on it

If any of those are true, get in touch. Bring the Gem — a working prototype makes the conversation much shorter.