Lesson 12 · ChatGPT Mastery Pro+ ~14 min Build + publish

Custom GPTs: agents in your voice.

A Custom GPT packages a system prompt, knowledge files, and tool actions into a reusable agent. Done well, it saves hours a week. Done badly, it's just ChatGPT in a costume. This is the build pattern that produces GPTs you actually keep using — starting with the part that makes or breaks them: the system prompt.

The mental model

Most GPTs fail at the system prompt.

The usual flop is a two-line prompt — "You're a helpful email assistant" — that doesn't actually change behavior. The fix is structure. A strong system prompt has four sections, each doing a distinct job: Identity (who it is and its one job), Voice (how it writes), Process (the steps it follows), and Don'ts (the hard boundaries). Build one now.

Do it · assemble the system prompt

Drop each line into its section.

Here are eight lines for an "Email Drafter" GPT. Send each one to the section it belongs in and watch the prompt come together.

Lines to place

System prompt · Email Drafter

## Identity

— empty —

## Voice

— empty —

## Process

— empty —

## What you don't do

— empty —

That's a system prompt that changes behavior. Identity gives it one job, Voice makes it sound like you, Process gives it steps to follow, and Don'ts draws the hard lines. Vague two-liners can't do any of that.

Your call · file or prompt?

Knowledge file, or system prompt?

Custom GPTs take up to 20 knowledge files — searchable, but not "memorized." Behavior belongs in the prompt; reference data belongs in files. Place each one.

Actions · do, don't just say

Auto-run, or require confirmation?

Actions let a GPT call real APIs. Read-only lookups can run freely; anything that creates, sends, or changes data needs you to confirm first.

Hard rule

Never wire an action with write or delete permissions to a production system without confirmation in the prompt. GPTs occasionally call actions you didn't expect — "show me what you'd do, then wait for 'go'" is the safeguard.

Build a GPT you'll use weekly

Pick one task you do 5+ times a week. Write its system prompt with the four sections, add 2–3 knowledge files if it needs reference data, and wire an action only if it truly needs to do something. Test the edge cases — off-topic prompts, ambiguous requests — then use it daily for a week.

Try a prompt in the playground → Back to ChatGPT Mastery

What you can do now

Write a system prompt with all four sections: Identity, Voice, Process, Don'ts
Put behavior in the prompt and reference data in knowledge files
Remember knowledge files are searched, not memorized — verify with test prompts
Add actions only when the GPT needs to do, not just say
Require confirmation for any action that creates, sends, or modifies

Pro+

Up next in ChatGPT Mastery

Lesson 13 · Code Interpreter — data analysis without writing Python

Upload a CSV or Excel file and let ChatGPT run real Python on it — cleaning, analysis, and charts, no code required. The final lesson in the track. Start lesson 13 →