Custom GPTs: build your first in 20 min.
A Custom GPT is a purpose-built assistant with its own instructions, conversation starters, and knowledge — your setup, baked in once. In August 2026 OpenAI closed new GPT creation on personal accounts, so where you build one has changed. What hasn't changed is the skill this lesson teaches, and it's the one that actually matters: writing instructions that constrain without paralyzing. Read the box below first, then build all three.
New GPT creation and publishing are no longer available on personal ChatGPT accounts — that's Free, Go, Plus and Pro. OpenAI's help documentation now states it plainly. Creation and publishing live in Business, Enterprise and Edu workspaces, on the web, where workspace settings and permissions allow it.
What still works on a personal plan: existing GPTs remain usable, and you can keep editing a GPT you already created if your plan and permissions allow. (Duplicating one doesn't — that needs create eligibility.) What to use instead: a Project with standing instructions and files does most of what people built GPTs for.
So read the three builds below as instruction-design practice. Every field you fill in here has a direct equivalent in Project instructions, custom instructions, and workspace GPTs — the box changes, the craft doesn't.
OpenAI is retiring Custom GPTs and migrating them to plugins. Per the Help Center’s retirement FAQ, the published Enterprise timeline (dates subject to change) is: Sep 11 admin notice · Sep 17 target for the in-product migration button and user banner · Sep 25 creation of new GPTs ends (publish any draft you want to migrate before then) · Dec 11, 2026 scheduled retirement, when Custom GPTs stop running and leave the directory. OpenAI says other plans “may follow the same transition timeline” with in-product notices, and the change affects all ChatGPT plans.
What migrates: the GPT’s instructions become a skill inside the new plugin and connected apps come along as apps. What doesn’t: custom actions (rebuild them with a connector or MCP server), the GPT’s chosen model, and possibly conversation starters and past chats. Only the creator or a workspace admin can migrate, the GPT must be published, the original goes read-only afterwards, and the replacement plugin starts private — share it deliberately. What to do now: list the GPTs you rely on, note who built each, save the instructions and reference files, and test the migrated plugin on a few familiar prompts before pointing your team at it. The plugin side is covered in Connectors and plugins.
A Custom GPT is your setup, baked in once.
Every time you open a blank chat and re-paste the same role, rules, and format, you're paying the setup tax again. A Custom GPT bakes that setup in permanently. You open it, type a short query, and get output already shaped to your patterns — no re-prompting ChatGPT from scratch each time.
The 20-minute build compounds across every future use. But the value lives entirely in the instructions — and there's a sweet spot most people miss.
Tune the instructions to the sweet spot.
You're building Sales Follow-Up Pro. For each line, pick the version that constrains without paralyzing — not so loose it's useless, not so rigid it can't write. Land all three and the GPT comes together.
Same pattern, different jobs.
Every build follows the same shape: a clear role, a structured output, calibrated constraints. Here are three worth copying.
Sales Follow-Up Pro
Drafts follow-ups using the specific-friction pattern and your tone.
Meeting Prep
Give it a topic and attendees; walk in prepared.
Exec Summary
Turns long content into a summary that respects a busy reader's attention.
Stored instructions, or this chat?
For each piece of context, decide whether it belongs in the GPT's saved Instructions or only in the live Conversation.
Build the one that pays off.
Build the GPT you'll actually use
Pick the workflow you run most often in ChatGPT. Build a Custom GPT for it in 20 minutes — name, calibrated instructions, 2-3 starters. Use it for the next week instead of regular ChatGPT. Most people feel the quality jump immediately.
What you can do now
- Build a Custom GPT in under 20 minutes with the four essential fields
- Write instructions that constrain output without paralyzing the model
- Decide what to keep in instructions (process) vs. each conversation (specifics)
- Reuse the build pattern for sales, meeting prep, and executive summaries
- Test a GPT against real use cases before you rely on it