Skip to content
Start small
Early preview Nothing here is released yet. Pages marked Garden are drafts that no person has reviewed. How review works

Custom assistants

Garden · not yet reviewed Established: widely used Widely discussed since 2023-11

A version of an AI tool set up for one purpose, with its own instructions, files and sometimes tools, that you or others can open and use without setting it up again.

Also called: Custom GPTs · GPTs · Gems · bots · custom agents · assistants

Connections

You may know it from

  • human resources a role with a job description Defining a role once (purpose, responsibilities, reference material) lets different people step into it.

In products

The same idea, under different names in different tools.

Provider What they call it
OpenAI Custom GPTs
Google Gemini Gems
Anthropic Claude Projects (shared)
Microsoft Copilot agents

In plain words

A custom assistant is an AI you have pre-configured for a single job. You give it a name, a set of instructions, some reference files and maybe access to tools, and then anyone with access can use it for that job straight away. OpenAI calls these GPTs, Google calls them Gems, and a shared Claude Project plays a similar role.

They are handy for repeated tasks with a clear shape: an onboarding helper loaded with the staff handbook, a feedback assistant that knows your grading rubric, a style checker for your brand voice. They also let one person’s setup help a whole team.

The weak points are the usual ones for any shared tool. Instructions drift out of date, files go stale, and people may trust its answers more because it looks official. Someone should own it and check it now and then.

Example

A school administrator builds a Gem loaded with the parent handbook and the instruction “answer only from the handbook; say when it doesn’t cover something.” Front-office staff use it to answer routine parent emails faster.

How it connects

  • Builds on custom instructions and the file-holding idea behind projects.
  • Parallels skills: both package a way of working, at different sizes.
  • Shared assistants need evals or spot checks to stay trustworthy.

Try it

Write the one-paragraph job description your assistant would have: purpose, audience, what it must never do. Paste it into a new chat as a first test, before you build anything.


Read this card as plain Markdown: /concepts/con-custom-assistants.md. Know a parallel from your own field, or a product that implements this? Add it.