Out-Of-Character Guidance
Use double parentheses when you need to guide the chat instead of speaking in-scene. The character should use the note silently, not respond to it as dialogue.
Chat Guidance
CO3 chats work best when the user can steer the scene without breaking the character, losing momentum, or fighting the model. These tools help you correct direction, continue a scene, test alternate replies, and report what feels off.
Empty Sends
Sending with an empty reply asks the character to continue the scene. This is useful when you want a beat, interruption, observation, or natural next moment without adding user dialogue or moving your own persona by hand.
Regenerations
Regenerate when a response is close but not quite right. It creates an alternate version so you can choose the reply that best fits the scene.
Saved Chats
A character can have more than one chat. Continue Latest Chat resumes the most recent saved conversation, while starting a new chat should keep the older one in history instead of replacing it.
Using Out-Of-Character Notes
Out-of-character notes are for corrections, style nudges, continuity reminders, or small steering instructions. Keep them direct and practical, and phrase them like guidance for future replies rather than dialogue the character can hear.
- Good: ((stop calling them Suns; Rocky calls them Stars.))
- Good: ((keep this in limited third person from the character's point of view.))
- Good: ((less reassurance here; let the silence feel awkward.))
- Good: ((the scene has moved to the apartment; stop describing the rooftop.))
- Avoid using OOC notes as normal dialogue the character should hear.
OOC notes are carried as recent chat guidance, not permanent character edits. If the same correction should always apply, update the persona card or send the note again when the chat starts drifting.
When To Empty-Send
Empty sends are for moments where the scene needs to breathe. The character might notice something, ask a question, react physically, shift the topic, or let the environment move forward. Empty sends should not give the character permission to control your user persona. If an empty send only restates the last beat instead of progressing the scene, flag it as a response-quality issue.
When To Regenerate
- Use regenerate when the response is repetitive, too flat, too safe, or too direct.
- Use it when the character missed point of view, scene location, revealed information, or emotional pacing.
- Pick the variant that best preserves the story path you want.
- Flag the issue instead if the response shows a recurring problem the platform should learn from.
Writing As And Time Passing
The chat remembers which user persona you are writing as for that chat. If you switch to another scene or another persona, check the Writing As selector before replying. After a longer real-world break or an in-story jump like "the next morning," CO3 may ask whether to continue exactly where the scene stopped or allow time-lapse effects.
- Continue exactly where you left off when no time has passed in the scene.
- Allow time-lapse effects when mood, sleep, routines, or offscreen consequences should matter.
- Report it if the chat invents major user actions, skips too far, or ignores the time jump you wrote.
Continuing Or Starting Fresh
Use Continue Latest Chat when you want to return to the newest saved conversation with a character. Start a separate new chat when you want a different opening, premise, user persona, or test path without overwriting the old conversation.
Rewind
Rewind removes later messages and returns the chat to an earlier point. Use it when a scene went in the wrong direction and regeneration is not enough. CO3 asks for confirmation because rewind cannot be undone.
Feedback In Chat
The Good Response and Flag Issue tools help identify what is working and what breaks immersion. A good report explains what felt wrong, not only that something was bad.
- Use Good Response when the character voice, pacing, tension, or emotional continuity feels right.
- Use Flag Issue when the response repeats itself, contradicts the scene, controls the user, loses point of view, breaks quotation marks, or feels generic.
- Add details when you can: what repeated, what was contradicted, what the character should or should not know, or where the scene actually is.
- Flag style patterns too: echoing the user's words, starting too many replies the same way, overexplaining body language, ignoring a subject change, or forcing one question too hard.
- Do not include API keys, passwords, private tokens, or billing details in feedback.
What To Test In Longer Chats
Longer scenes are where CO3's chat behavior matters most. Useful notes include whether the character remembers the current location, respects what has or has not been revealed, lets some questions drop, carries emotional residue without repeating it every turn, and keeps their own voice instead of sounding like every other character.
What CO3 Is Trying To Preserve
The goal is collaborative story movement without stealing the user's agency. The character can introduce observations, tension, questions, hesitation, interruptions, and environmental changes. The character should not decide what you think, feel, say, choose, forgive, reveal, or want unless you wrote it first.