---
title: "What your twin answers"
description: "Keep it on your subjects instead of behaving like a general chatbot."
source: https://hiy.ai/docs/what-it-answers
---

# What your twin answers

Left alone, a language model will answer anything — including things you'd never speak on. These settings keep your twin answering *as you*, on *your* subjects. They're in **Voice & rules → What it answers**.

## What you cover

List your subjects: "pricing, positioning, early-stage fundraising". Leave it empty and your twin works it out from your profile and your sources — which is usually fine, but saying it explicitly is more precise.

## How firmly it declines

The **Guardrails** tab reads this setting back in one line — *"Asked something
outside that: …"* — and its **How this works** link lands here. The subjects the
setting applies to are [What you cover](#what-you-cover), just above.

Three settings, in order of strictness. Beside each is the value stored in
`scope_strictness` — what your account export and the API call it — so a
support reply can name one exactly.

Which one a twin *starts* on depends on the kind of twin. Your own twin starts
on **Balanced**. [Support Agents](/docs/support-agents) and
[Team Agents](/docs/team-agents) start on **Strict**, because a confident guess
about your product or a company matter is worse than an honest "I don't have
that". It's a starting point either way — change it whenever you like, and
changing the kind of twin never changes a setting you already chose.

- **Strict** (`value: strict`) — Only answers what your own material supports. Anything else gets an honest "I don't have that". Choose this when being wrong is worse than being unhelpful.
- **Balanced** (`value: balanced`) — Answers within your subjects, and warmly declines unrelated things: trivia, coding help, homework, current events. It offers what it *can* help with instead of leaving a dead end.
- **Loose** (`value: loose`) — Stays in your voice and leads with your material, but will help with adjacent general questions. Still declines anything clearly unrelated.

## Off-limits topics

Subjects your twin refuses or deflects regardless of everything above — family, client names, anything private. Your personality interview can add to this list too — on a personal twin or a [Team Agent](/docs/team-agents). A [Support Agent](/docs/support-agents) has no interview, so its off-limits list is only ever what you type, and the hint under its editor says so.

**The list holds ten.** That is the ceiling in the editor, in the API and in
what the interview may add, and it is now one number rather than three: the
interview used to merge in up to twenty boundaries, which wrote rows the
settings page then refused to save. If an interview states more boundaries than
there is room for, the ones already on your list are kept — they are what *you*
typed — and the screen at the end of the interview says how many did not fit
and what to do about it ("2 boundaries didn’t fit — your off-limits list is
full at 10. Remove a topic in Voice & rules → What it answers and state it
again."). If a list from before this change is still over ten, the group says
so and how many to remove, and everything else on the page saves normally in
the meantime.

**Note:** Off-limits is not the same setting as scope, and the two combine rather than
compete. Scope is *what you cover*; off-limits is *what you won't discuss even
though it is yours*. The off-limits list is added on top of whatever strictness
you chose, so widening your scope never re-opens one of these subjects.

## Saving what you change

Voice & rules is one form with one **Save changes**, and it lives on a strip
that appears at the bottom of the page the moment anything is unsaved. The strip
names the groups holding an edit — "Unsaved changes in What it answers" — so a
change made in a group you have since collapsed cannot be forgotten. **Discard**
beside it puts every field back to its last saved value and asks nothing first,
because everything it drops is still on screen in front of you.

If a save is refused, the reason is printed **on the strip**, beside the button
that produced it — once, and nowhere else. A confirmed **Saved.** appears under
the last group, and only while nothing is pending: it is never on screen at the
same time as "Unsaved changes in …".

**Leaving the page asks first.** Click something else in the sidebar, switch
agents, open the agents list, or press your browser's Back button with an edit
still pending, and hiy stops to ask — *Leave without saving?* — naming the same
groups the strip names, over **Stay on this page** and **Discard and leave**.
There is deliberately no Save on that question: the strip behind it already has
one, and Stay puts you back in front of it. Escape and a click outside both
mean Stay, so nothing you can dismiss by accident throws the edit away. hiy
keeps no draft, which is why it asks rather than tidying up after you.

The same question guards the other two pages that track an unsaved edit —
[your Profile](/docs/profile#saving-what-you-change) and, on a Support or Team
Agent, [Escalation](/docs/escalation#saving-your-changes). **Closing the tab or
reloading is your browser's question, not ours**: hiy asks it to warn you, but
the words are the browser's own and no group is named in them.

Everything on this page is set here, in **Voice & rules**, and read back on
one screen: the **Guardrails** tab of
[Agent knowledge](/docs/knowledge-library#guardrails) lists what an agent
refuses, what it covers and what it does when asked something outside that —
in full, without opening a form.

## A note on gaps

Declining isn't failure. The clearest of those declines become questions in your queue — hiy only queues the ones it can be sure were an admission, so what you get is a list worth working through rather than every answer that happened to hedge. Answering one teaches your twin permanently, and that loop is why the honest setting is the useful one.