---
title: Assistants
description: An assistant is a saved set of standing instructions you can talk to. What you can configure, what each setting actually changes, and when building one is worth the two minutes.
---

An assistant is a named set of standing instructions — how it should behave, which of
your apps it may touch, which engine it runs on, and what it should ask you before
spending. You talk to an assistant instead of re-typing the same briefing at the top of
every chat.

You already have one. The **Assistants** screen opens with **Everyday assistant**,
tagged **DEFAULT**, and the screen states the deal in a line: *"One helper for
everything — or build your own for the tasks you do often. Give it a personality,
abilities, your apps and your files. No technical skills needed."*

## Assistant, or a thing you taught it?

They solve different problems and they do not overlap.

- An **assistant** is *who you are talking to*. It carries instructions, files, an app
  scope, an engine and a budget, all at once. You pick it, then you chat.
- A **thing you taught it** — a personal Add-on — is *one reusable expectation*: a
  format, a checklist, a house rule. It lives on your account and is available in any
  chat, whichever assistant you are talking to. It is not attached to one of them.
  See [Add-ons](/docs/add-ons).

So: teach it a rule once when the rule should apply everywhere. Build an assistant when
a whole *posture* recurs — the standing brief, the file it should always consult, the
narrow set of accounts it should be allowed near.

## Building one

**Create an assistant** — *"takes about two minutes"* — opens a builder with five
fields. Only two are required.

**Name and mark.** A name up to 60 characters, and one of four colours for its tile.

**How it should behave.** Up to 2,000 characters, and this is the field that does the
work. Write it the way you would brief a new hire, in your own words: *"You plan trips
with realistic budgets. Always check real prices and opening days. I prefer trains over
flights under four hours."* Whatever you write here is given to the assistant at the
start of every chat you have with it.

**What it knows — add your files.** Up to five files, **500 KB each**. One honest
detail worth knowing, because it changes what you should expect: when you hand this
assistant a **task**, the text of those files is given to it in full. In an ordinary
**chat**, it is told the file *names* but not their contents. If the file content is the
point, hand off a task, or attach the file to the message — see
[Giving it materials](/docs/giving-it-materials).

**Which of your apps it may use.** This narrows the assistant to a subset of your
connected [Apps](/docs/connecting-apps). Pick nothing and it keeps the full set available;
pick Gmail alone and it cannot reach the others, even when they are connected. Note that
the picker currently offers only three of the six first-class apps, so if you want an
assistant that can use Slack, GitHub or Airtable, leave this field empty rather than
choosing from the short list.

**Engine and budget.** The engine this assistant defaults to — **Auto**, **Fast**,
**Smart** or **Max** — and the point at which it should stop and check with you, offered
as **Ask above $0.10**, **$0.50**, **$1.00** or **$5.00**. Both are defaults it carries
into new work, not locks: you can still change the engine on an individual conversation,
and a task can be given its own limit. The budget figure is a threshold you choose, not
a price. What a run costs depends on the engine it runs on — see
[Spending limits](/docs/spending-limits) and [Model prices](/docs/model-prices).

## What the card tells you afterwards

Each assistant's tile shows **how many tasks it has done** and the **average cost** of
those tasks. Both numbers are counted from your own history: the tasks filed under that
assistant, and what was actually charged against them. They are a record, not an
estimate — a brand-new assistant honestly reads zero.

## Editing, and what you cannot edit

Your own assistants can be edited or deleted at any time from the pencil on the tile;
deleting one removes its instructions and its attached files permanently, and the
confirmation says so. **Everyday assistant** has no edit control, because it is not a
row you own — it is the built-in default, and the server refuses to change it.

An assistant is private to your account. There is no sharing, no publishing, and no
common pool.

## When it is worth it

Build one when you have caught yourself re-typing the same three sentences of context.
Build one when the safest version of a job is a *narrow* one — an assistant scoped to a
single app is a smaller blast radius than a general chat. Build one when the standing
budget for a kind of work is different from your usual.

Otherwise, keep using **Everyday assistant**. Its whole job is to be good enough that
you do not have to configure anything, and a single well-written message beats a badly
configured assistant every time — see [Writing a good request](/docs/writing-a-good-request).