Custom GPTs: using and judging ready-made assistants
Understand what custom GPTs are, how to choose and test them, how they are built, and how to stay safe with third-party tools.
In this lesson you will
- Explain what a custom GPT is
- Choose and test one safely
- Read permission prompts carefully
- Recognise good GPT instructions
About 15 minutes
A GPT is a pre-packaged assistant
Hira finds a “Resume Coach” in the GPT store and a “Spanish Tutor” that a teacher made. She opens one and immediately gets an assistant that already knows its job: it greets her with the right questions and keeps to a clear format. She wonders: how is that different from the regular ChatGPT? And can she trust it? In this lesson you will learn what custom GPTs are, how they are built, how to choose and judge them, and what to watch for.
Features around custom GPTs and any store change often. Who can use them, create them or publish them depends on your plan and region. Check OpenAI’s current help pages.
What is a custom GPT?
A custom GPT is a version of ChatGPT that someone has set up for a particular purpose. The creator gives it:
- Instructions: a long, careful prompt describing its role, rules and style.
- Knowledge files (optional): documents it can draw on.
- Capabilities: such as web search, image creation or data analysis.
- Actions (optional): connections to outside services through an API.
- A name, picture and conversation starters so it feels like a ready-made tool.
Think of it as a saved project with a professional front door. It is not a different brain; it is the same underlying technology dressed and instructed for one job.
Why use one?
| Benefit | Example |
|---|---|
| Saves prompt writing | A GPT that always asks about your audience before writing an ad |
| Consistent format | A meeting-notes GPT that always outputs Decisions, Actions, Questions |
| Specialised knowledge | A GPT built around a company’s policies or a course’s material |
| Sharing with others | A teacher shares a study GPT with a class |
| Guided experience | Conversation starters walk beginners through tasks |
Finding and choosing a good one
- Search by task (“CV review”, “recipe planner”) in the GPT section or store.
- Check the creator: a known organisation or named expert beats an anonymous account.
- Read the description and ratings, but treat them as hints, not proof.
- Try it with a test question you already know the answer to.
- See what capabilities it uses (web, files, actions), and what it asks of you.
- Compare it with a plain ChatGPT chat using your own good prompt. Sometimes the plain one is just as good.
Safety and trust
- Your chats may go to the creator’s services. A GPT with actions can send information to an outside service. Read the permission prompts before you allow anything.
- Do not share sensitive data with a GPT you do not fully trust.
- Names can mislead. A GPT called “Doctor” or “Tax Advisor” is not a doctor or a tax adviser.
- Look for fake or copycat tools using famous brand names without being official.
- Check claims exactly as you would from any chatbot (lesson 7).
- Keep your own judgement. A confident persona is still a text generator.
Giving a third-party GPT permission to connect to your accounts, email or files without reading what it will do. Approve only what you understand and need, and remove connections you no longer use.
A test routine for any GPT
- Ask something with a known answer. Does it get it right?
- Ask something it should refuse or say it does not know. Does it bluff?
- Ask it to explain how it works. What rules does it follow?
- Try one real task and check the result against your own standards.
- Note whether it asks for more information than it needs.
Before we start, tell me what you can and cannot do, what data you will use from this chat, and whether you send anything to external services.
A trustworthy GPT should give a straightforward answer. Remember that its own description of itself is not proof, so combine this with the permissions you can see in the interface.
How custom GPTs are built (concept)
If you can create GPTs on your plan, the builder usually lets you describe what you want in plain language and then tune the instructions. The strongest GPTs share these features:
Role: You are a friendly CV coach for first-time job seekers in Pakistan.
Process:
1. Ask for the target job and the person's experience, one question at a time.
2. Rewrite each experience as bullet points starting with an action verb.
3. Never invent achievements. If information is missing, ask.
4. End with a checklist of three improvements.
Format: Short sentences. British spelling. Bullet points for lists.
Boundaries: Do not give legal or immigration advice. Do not ask for ID numbers or addresses.
Notice that it looks just like a good project instruction: role, process, format, boundaries. If you cannot create a custom GPT, you can build almost the same experience with a project (lesson 12) or by saving the instructions in a note and pasting them at the start of a chat.
Testing and improving instructions
- Write five typical requests and five tricky ones, including attempts to make it break its rules.
- Run them all and note failures.
- Fix the instructions, run again.
- Ask a friend to try it without any hints.
- Avoid very long instructions. Clear rules in priority order work better than essays.
Privacy for builders
- Do not upload confidential documents to a GPT you share with others, because knowledge files can sometimes be extracted by clever users.
- Do not include secrets or API keys in instructions.
- Say plainly in the description what the GPT does and does not do.
- Respect copyright: upload only material you may share.
When a GPT is better than a plain chat
| Use a GPT when… | Use a plain chat when… |
|---|---|
| The task is repeated often | It is a one-off |
| You need a consistent format | The format changes each time |
| Others need an easy start | You are experimenting |
| You trust the creator | You are unsure about a tool’s source |
| It uses documents or actions you need | A good prompt does the job |
You would not install every app on your phone without a glance at who made it. Give custom GPTs the same quick look before you trust them with real information.
Check what you learned
Five quick questions. Pick an answer and you will see straight away whether it is right and why.
What is a custom GPT?
- A new computer
- A virus
- A paid camera
- A ChatGPT version set up for one purpose
It has instructions, optional files and capabilities.
How should you test a new GPT?
- Ask something where you know the answer
- Give it your passwords
- Never test
- Ask for jokes only
A known-answer test shows reliability.
What is a risk of GPT actions?
- The screen turns off
- Longer names
- Data may go to outside services
- More colours
Read permissions before approving.
A GPT named ‘Doctor’ means…
- It is a real doctor
- It is free
- It is private
- Nothing about real qualifications
Names do not prove expertise.
What often works as well as a custom GPT?
- A blank chat
- A project with good instructions
- No prompt
- A long email
Projects or pasted instructions can replicate it.
Summary
- A custom GPT is ChatGPT set up for one job.
- Check the creator, test with known answers and read permissions.
- Do not share sensitive data with tools you do not trust.
- Projects or saved prompts can often do the same job.
Questions
Can I make money from a GPT?
Programs and rules for this change over time. Check OpenAI's current policies, and respect copyright and platform rules.
Are GPTs safer than plain chats?
Not necessarily. They add the creator's instructions and sometimes outside connections, so trust matters more.
Can I delete a GPT chat?
Yes. Chats with GPTs are usually part of your history and can be deleted like others, but check how the creator handles data.
