How to Give a Custom GPT a Personality (Without It Drifting)

You build a custom GPT, give it a name, and type "be friendly and helpful" in the instructions box. A week later it reads exactly like every other GPT.

A custom GPT personality isn't a tone setting. It's a set of written behaviors the model can follow every time. Here's how to write one that holds up.

Where the Personality Lives

In the GPT builder, open the Configure tab. The Instructions field is where the personality goes. The conversation starters and description help users, but they don't shape behavior much. The instructions do almost all the work.

Uploaded knowledge files are for reference material, not personality. Keep the character in the instructions field.

Step 1: Pick One Clear Job

A GPT that does everything has no personality, because there's nothing for the personality to attach to. Choose one job: inbox triage, product spec reviews, study sessions. The character grows out of the job.

Step 2: Write the Character Brief in Two Sentences

Before writing instructions, answer two questions for yourself:

• What does this character care about most?
• What does it do that a generic assistant wouldn't?

For an executive assistant GPT, the answer might be "cares about protecting your time, and will tell you when your calendar is the problem." That second half is the personality.

Step 3: Turn Traits Into Behaviors

"Be organized" means nothing to the model. "End every reply with a short list of next actions, each with an owner" is something it can do. Go through each trait and rewrite it as an observable behavior.

Step 4: Set the Formatting Rules

GPTs drift most on format. Lock it down:

• Reply length (for example, under 150 words unless asked for more)
• Structure (summary first, then details)
• What it never does (no long preambles, no restating the question)

Step 5: Add the Hard Cases

Tell the GPT what to do when a request is vague, out of scope, or when the user pushes back. These instructions are what keep the character intact when the conversation gets off script.

Step 6: Test in Preview, Then Revise

Use the Preview pane on the right of the builder. Send a normal request, an off topic request, and a vague one. Revise the instructions wherever the character slips, then test again.

Example Custom GPT Instructions

Here's a starting point for an executive assistant GPT:

## Role
You are an executive assistant for a busy founder. Your job is
to protect their time and keep commitments from slipping.

## How you write
- Open with what you did or what you need, in one line.
- Keep replies under 150 words unless asked for more.
- End with "Next:" and a short list of actions with owners.
- Warm but direct. No filler, no long apologies.

## What you care about
- Time: point out meetings that could be an email.
- Follow through: track anything the user says they will do.
- Clarity: if a request is vague, ask one specific question.

## Hard cases
- Out of scope (legal, medical, tax): say so and suggest
  the right kind of professional.
- User disagrees: restate their priority, adjust the plan,
  and note any tradeoff in one sentence.
- Missing information: list exactly what you need.

Markdown headings work well in the GPT instructions field. They keep the sections easy to scan when you come back to edit.

Or Start From a Finished Persona

If you'd rather skip the drafting, CastMyAgent has complete persona packages you can paste straight into the GPT builder. They also work with Claude and Gemini.

Val, the executive assistant, handles the task and then tells you, kindly, what needs fixing on your calendar. Zara, the AI product manager, won't move forward on a feature until you can write its headline. Each package includes the production prompt, a character brief, a style configuration, and behavioral encoding.

All 19 walk-ons are free. Browse the roster at castmyagent.ai.