Set your Persona and Custom Instructions

Review scope: Linked public documentation checked October 30, 2026. Examples are proposed checks, not recorded OD results; use your account's current controls.

You will tell the chat who you are playing, then give it a small set of writing instructions.

A Persona describes your role in the conversation. Custom Instructions describe how you want replies written. Mira's personality belongs in Mira's character definition; it does not belong in your Persona.

Before you begin

Have an account and a practice chat. Use a separate practice Persona if you already have one you want to preserve.

The Personas and instructions guide, last modified July 27 and re-read November 1, 2026, documents a Persona description allowance of 2,500 characters and a Custom Instructions allowance of 15,000 characters. Check the current field counter in your account. These are allowances, not writing targets or a specification of the model's total context; keep both fields focused rather than filling the available space.

Step by step

1. Open your profile settings. Find the Persona controls. Create a new Persona using the control available there, or edit the practice Persona you intend to use.
2. Give it a name. For this example, use Alex. Choose the other identity fields according to the role you want to play.
3. Enter a short description. Use the next box as the Persona description.

Alex is an adult traveler visiting the town archive for the first time.
Alex carries a worn satchel and is interested in old maps. Alex has no
established friendship with Mira and has not yet chosen how to investigate.

4. Save the Persona. Open the intended chat and select Alex using its Persona control. Creating a Persona and selecting it for a chat are separate actions. Reopen the selector to confirm the saved selection before testing; keep a copy of any existing Persona or instructions you replace.
5. Check image inclusion. If Include persona in images is offered, choose whether Alex's appearance should be included. Selecting that option can affect who appears in a requested scene; it does not turn the written Persona into a reference photograph.
6. Open the chat settings. Find Custom Instructions. Check whether the interface applies the text to this chat or a broader scope before replacing any existing instructions.
7. Add a small instruction set. Paste the next box into Custom Instructions, not into the Persona description.

Describe the environment and the AI characters' actions and dialogue.
Leave Alex's words, actions, decisions, and private thoughts to me.
Keep each character's knowledge separate.
End at a useful point for me to respond.

8. Test with a new message. Ask Mira: “What would you suggest I investigate first?” Look for an answer that offers a useful possibility without deciding that Alex accepts it.
9. Correct only the relevant layer. If the wrong person is being described, check Persona selection. If the AI is writing your choices, revise the behavioral instruction. If Mira has the wrong occupation, correct Mira's character fields.

The Known State Personas and instructions guide distinguishes user identity, persistent response rules, and temporary scene direction.

What belongs where

Use chat controls for temporary directions rather than keeping a one-time arrival in a persistent field.

Make one rule specific enough to inspect

For an original Custom Instructions practice revision, replace make this immersive with the behavior you actually want:

When Mira answers a practical question, address that question through her
dialogue or an observable action. Use details relevant to the archive scene.
If a fact is not established, leave it uncertain or ask for the missing fact.
Leave Alex's reply and next choice to me.

Keep only the parts your chat needs and reconcile equivalent rules already present. Try an ordinary catalog question, then an uncertain address question. Judge relevance, uncertainty, and user agency separately. The OD prompt lab explains how to compare a revision without changing the whole setup.

Request the chat language deliberately

OD's language help article, edited November 5, 2025 and read October 31, 2026, says a language request may take several prompts to take effect. This describes chat behavior; it does not establish equal fluency across languages or change the interface language.

For a continuing preference, try this original Custom Instructions example after saving the existing text:

Write dialogue and narration in Spanish. Keep the established names,
character roles, and story facts. Leave my actions and decisions to me.

Ask Mira, where should a visitor begin looking for a map? and inspect several ordinary replies. Record language use and preservation of facts separately. Remove an older conflicting language request before judging the result. If you want a single translated passage instead, request that passage in the current message rather than changing the whole character.

Check your result

Your selected Persona should match the role you want for this conversation. Mira should still have her own identity. Judge the actual replies: an instruction states the intended behavior but does not guarantee every response.

If something goes wrong

• The AI uses a different name: inspect the selected Persona and old memory entries for conflicting names.
• Every reply repeats an event: remove that event from persistent instructions and give it once as a scene direction.
• The AI obeys one rule while breaking another: look for contradictions such as “never advance the scene” alongside “always introduce a new event.”

Next: Memory and continuity, or chat controls and slash commands.

Back to the tutorial index

Optional non-graphic dialogue guideline

The voice, pacing and consent module provides a downloadable text block and placement steps. Open the file and copy its contents into the intended chat's Custom Instructions. A supplied May 16, 2026 screenshot places this field under Memory; inspect the controls and Current Chat/Default scope actually shown in your account before saving. The four proposed checks have no recorded result for this module.