Most people use AI assistants the same way every time: open a new chat, explain the context, paste in some reference material, then ask the question. If you do a particular task regularly, such as drafting client emails, reviewing proposals or writing product copy, you can save that setup and reuse it.
Updated September 2026: we added links to each company’s documentation for custom assistants and to the research on learning from examples.
The main assistants each offer a version of this:
- ChatGPT: Custom GPTs and Projects
- Claude: Projects, with custom instructions and knowledge files
- Gemini: Gems
The names and details differ, but the ingredients are the same. Google’s tips for Gems, for example, break instructions into persona, task, context and format, and let you add files for reference.
Ingredient 1: Clear instructions
Instructions tell the assistant who it is working for and how to behave. Good instructions cover:
- The role and audience: “You help our support team draft replies to customers of a home insurance company.”
- The goal: “Replies should resolve the question in as few messages as possible.”
- Style: tone, length, formatting, words to avoid.
- Rules: “Never promise a claim outcome. If a policy question is not covered by the attached documents, say you will check with a specialist.”
- Process: “Ask for the policy number before answering account-specific questions.”
Write them like a brief for a new colleague. Our prompting guide covers the same principles.
Ingredient 2: Reference files
Upload the documents the assistant should rely on: style guides, product information, policies, templates and good examples of past work. The assistant retrieves from these when answering, which is the same idea as the RAG technique used in business chatbots.
Keep files current and remove outdated versions. Conflicting documents produce conflicting answers. Limits vary by tool; a custom GPT, for example, accepts up to 20 files.
Ingredient 3: Examples
One or two examples of ideal output are often worth more than a page of instructions. Learning from a few examples in the prompt was one of the key findings of the GPT-3 research in 2020. Include an example input and the response you would want.
Starter template You are [role] helping [audience] with [task]. Always [key rules]. Never [prohibited actions]. Use the attached [documents] as your source of truth, and say so when something is not covered. Write in a [tone] style, usually under [length]. Here is an example of a great response: [example].
Test before you rely on it
- Try five to ten realistic requests, including awkward ones.
- Try a question the files do not cover; the assistant should say so rather than invent an answer.
- Check that it follows formatting and tone rules consistently.
- Refine the instructions based on what went wrong.
Sharing with a team
On business plans, custom assistants can usually be shared within your organization, which spreads consistent quality across the team. Assign an owner to keep instructions and files updated, and follow your AI use policy about what data can be uploaded.
Good first candidates
- A brand voice writer for social posts and newsletters
- A proposal reviewer that checks drafts against a checklist
- A meeting-notes formatter that turns transcripts into your standard template
- A study tutor loaded with course notes
- A policy question assistant for internal HR or IT FAQs, with human review
Start with one task
A custom assistant takes perhaps half an hour to set up well and can save that time every week. Start with one task you repeat often, give it clear instructions, good files and examples, then test and refine it like any other tool.
Sources
- Creating and editing GPTs, OpenAI Help Center
- What are projects?, Claude Help Center
- Tips for creating custom Gems, Gemini Apps Help
- Language Models are Few-Shot Learners, arXiv, 2020
