How to roleplay with ChatGPT Play a story in ChatGPT, and keep it on track.

Narrator or character: pick your setup

Story roleplay works on any ChatGPT plan, and there are two ways to set it up:

This guide sets up a narrator, with notes for character play where it differs. OpenAI’s usage policies apply to stories as they do to any other chat.

Custom instructions or a project?

ChatGPT has two places for standing directions, and they suit different jobs:

• Custom instructions apply to every chat you have. OpenAI’s help center gives them 1,500 characters on the Free and Go plans and 5,000 on Plus, Pro, and other paid plans. Use them for how you like ChatGPT to talk to you in general.
• Projects gather chats, files, and instructions around one piece of work. A story’s project holds its rules in the project instructions, its notes as files, and every chat you play in it.

A project also lets you choose how memory works. With project-only memory, the project’s chats can draw on each other but not on your other conversations, and the story stays out of the memory ChatGPT uses everywhere else. That keeps the plot and your other chats apart.

Set it up once:

1. Create a project named after the story.
2. Set its memory to project-only.
3. Paste the brief from the next section into the project instructions.
4. Add a file called Story so far, even if it starts nearly empty.

A brief to paste into the project

Write the project instructions as a brief to a storyteller. Rules go near the top and backstory stays short; longer notes can live in the project’s files.

For one character instead of a narrator, replace the first line with the character’s own brief: who they are, how they talk, what they want from you, and a line or two of their dialogue to set the voice. Keep the rule about your character, and add that they speak only for themselves.

When ChatGPT plays your part

Sooner or later a reply decides something for your character: a line of dialogue, a nod, a change of heart. Each reply sets an example for the next, so deal with the first one:

• Edit and resend. Edit your last message, add a reminder, and send it again, so the faulty reply never becomes part of the story.
• Correct it in one line. [You wrote what Tamsin said. Rewrite your reply and stop before she speaks.]
• Tighten the ending. [End each reply on someone else’s action or question.]
• Check your own examples. If the opening or any sample dialogue has your character speaking, ChatGPT takes that as permission.

If it keeps happening late in a long chat, start a fresh chat in the project from your summary, as described below.

First scene, then one turn at a time

Open with a scene that’s already moving, and leave the first move to ChatGPT:

Then play one turn at a time. A turn says what your character tries or says, and leaves the result to the story:

• Tamsin hands over the letter, then asks the porter who paid for its delivery.
• She steps back onto the bridge, out of reach, and asks to see the dead man’s rooms.
• She breaks the seal and reads the letter on the glide down.
• [Skip to the next morning at the courier hall.]
• [Bring in a rival courier who wants this route.]
• [Keep the next scene quiet. No danger for a while.]

Mix in quiet turns, too. A question to a stranger, or a moment spent looking around, gives the world room to show what it holds.

Why details slip, and how to keep them

Every chatbot can read only so much of a conversation at once, so a long story eventually runs past what ChatGPT has in view, and early details start to slip. Memory helps within limits. OpenAI describes ChatGPT’s memory as a continually updated summary of context from your past chats, which suits knowing your preferences. A story’s chapter-by-chapter facts need a record of their own.

That record is the Story so far file. A routine that keeps it current:

1. At the end of each sitting, ask for a recap with the prompt below.
2. Replace the Story so far file in the project with the new version.
3. When replies forget names or contradict earlier scenes, start a new chat in the project and paste the resume prompt.
4. Keep a cast list file too: one line per recurring character, with how they relate to you.

To pick the story up in a new chat:

When the story outgrows the chat

ChatGPT is built to help with almost anything, and with a project and a recap it can keep a story going for weeks. Many players who want a story to run for months move it to LoreWeaver, where the record keeps itself while a storyteller runs the world and you play in it:

• Memories you don’t have to write. Every 4 replies, LoreWeaver saves what happened, and the memories that fit the scene come back on their own, found by meaning, wording, and who is present. A running recap covers the whole story, and you can edit it.
• Lore cards for the cast. People, places, and things each get a card that joins a reply when they’re named, or every reply once pinned. From Plus, Lorekeeper suggests cards for the characters the story invents, and the LoreMap shows how they’re linked.
• The Loom. Threads fire planned events when their conditions are met, and trackers show health, gold, quests, inventory, and dice rolls on screen.
• GPT models, built in. LoreWeaver runs them itself, with no ChatGPT plan or API key needed. GPT-6 Luna opens on Plus, and free accounts can play it with their 25 premium replies a day. GPT-6.1 Terra opens on Ultra. Claude and Gemini models are there too, among 43 in all, and you can change models mid-story without losing its memories or lore.

To move a story over, describe it in a sentence or two in the world builder, which writes the premise, the people and places as lore cards, who you play, and an opening. Paste your Story so far into the story’s recap, and carry on from the last scene. If you mostly talk with one character, the builder makes character worlds too. LoreWeaver accounts are for adults, 18 and over.

Sources

• OpenAI Help Center: ChatGPT custom instructions
• OpenAI Help Center: Memory FAQ
• OpenAI Help Center: Projects in ChatGPT
• OpenAI usage policies

All checked October 30, 2026.