How to stop an AI roleplay from speaking for your character

01Why the usual way fails

Language models are trained to continue text, so they keep writing past the point where the other character should stop. In a roleplay that means they often write the player's actions too. Prompt instructions like "never speak for the user" help for a while but fade in long sessions. Many chat apps also add their own system prompts and filters on top, which you cannot see or change. And when you bring your own model or key, you still have to clean up every reply by hand. Group scenes make it worse, because the model has to voice several characters and easily slips into yours.

02How The Codex AI helps

The Codex AI is a browser-based roleplay and story platform with a feature called Sovereign Guard. It cleans the model's output in several layers so the AI does not take over your actions, write fake lines for your character or add forced choices, and it lets you trim a reply with one click. You can use the free platform model or connect your own provider key or a local model. Your stories can also branch, so you can try another path without losing the first one.

03Step by step

1. Sign up and start on the Free tier with the platform model, or connect your own API key or local model.
2. Import a character card or create a new character and world.
3. Pick a session mode: Story RPG, Cast Lounge or Chronicle novel.
4. Play your turn and let Sovereign Guard remove lines the AI writes for your character.
5. Use one-click trimming for anything left, or branch the story to try a different choice.

04Tips

• End your turn with an action or a question, not an open sentence. A clear stopping point makes it easier for any model to answer only for its own character.
• Keep important world facts in the lore, not in the chat. When lore is pulled in by relevance on each turn, the model needs fewer reminders and drifts less.
• Branch instead of rerolling the whole scene. If a reply goes the wrong way, a branch keeps the good part of the story and lets you test another direction.
• Try a different model when one keeps puppeteering. Some models follow turn boundaries better than others, and being able to switch providers is the quickest test.

05Questions