A custom GPT is ChatGPT pre-loaded with your instructions, your files and a chosen set of tools. Read this before you start: new GPT creation and publishing are no longer available on personal ChatGPT accounts. You need a Business, Enterprise or Edu workspace, and your workspace settings have to permit it. On a personal Plus or Pro account, use ChatGPT Projects instead — same instructions-and-files idea, without sharing or actions.
1. Open the builder, then switch to Configure
The builder offers a conversational route where you describe what you want, and a configuration view where you edit the fields directly. Use Configure: it is the only place you see exactly what the GPT will be told.
2. Write instructions that constrain the answer
Instructions are a system prompt with a form around it. Vague instructions get ignored; rules with a shape get followed.
You are a support assistant for [product].
- Answer ONLY from the uploaded docs. If the answer isn't there, say so and point to support.
- Keep answers under 120 words, then offer to go deeper.
- Never invent prices, dates, or policy.
What it does: fixes the role, forbids answering outside the knowledge files, caps the length, and names the failure mode you most want avoided.
OpenAI’s own guidance: when the GPT must apply specific classifications, include short examples of acceptable and unacceptable output, and use headings and lists so priorities stay visible.
3. Add knowledge and capabilities
Knowledge takes up to 20 files, each up to 512 MB — reference material the GPT draws from during a conversation, as opposed to instructions, which say how to behave. Capabilities are the toggles for tools such as web search and image generation. Actions let the GPT call an external API you describe with an OpenAPI schema.
Verify it worked
Save, then open a fresh chat with the GPT and ask something deliberately outside the uploaded files. A correctly instructed GPT says it doesn’t know and points at support; one that confidently invents an answer needs a firmer instruction, not more files. That failure is ordinary hallucination — grounding it in the knowledge files is the fix. Next: write a system prompt that steers the model.
Source: Creating and editing GPTs.