How do you change perspective in Janitor AI?
Janitor AI has no point-of-view setting. The model copies whatever perspective is already in the visible text, so you change POV by editing it: rewrite the bot's greeting, your persona, and your last two or three messages in the perspective you want. It usually locks on within two turns.
Why — the first-principles explanation
There is no POV button anywhere in Janitor AI, and there never was one. That surprises people, but it follows from how the thing works. A language model does one job: look at the text in front of it and predict what comes next. It has no separate dial for tense, tone, or perspective. Perspective is just a pattern in the words, and the model reproduces patterns it sees. Janitor's own writing guide puts it bluntly: JLLM mirrors you. If your text says "You feel the cold," it keeps writing "you." If your text says "Mira felt the cold," it keeps writing "Mira."
This is why the greeting matters more than anything else you can change. Janitor's character-creation documentation says the initial message "sets the tone for the LLM as well, as most will try to follow the pattern of speech and actions dictated in the initial message" — and it recommends third person specifically because "third person often provides the best roleplay results in terms of preventing LLMs from speaking for the user." The greeting is the longest, cleanest, most prominent example of style in the whole chat. If it is written in second person, the bot will keep pushing you back into second person no matter how many times you ask it to stop.
That is also why typing "stop writing in second person" usually fails. An instruction is one sentence. The visible chat history might be forty messages of the opposite pattern. The evidence outweighs the order, every turn. Janitor's guide makes the same point about a related problem — vague "don't speak for me" prohibitions tend to backfire, and reframing the narrator's job works better than banning something. Tell it what to do, not what to avoid, and back the instruction up by changing the actual text.
One more mechanism worth knowing: editing is retroactive. The model has no memory of what your chat used to say. Every turn, it is handed the current text of the conversation and nothing else. So when you edit an old message, you are not correcting the record — you are rewriting the model's entire evidence base. Fix the greeting and your last few messages, and as far as the model is concerned, the chat was always in that perspective.
An example that makes it click
Think of the bot as a kid playing copycat. Whatever you do, it does back. You raise your left hand, it raises its left hand. Now you say, out loud, "stop copying me" — while still raising your left hand. The kid copies the hand. Your words are one small input; your behavior is the loud, obvious one.
So stop copying yourself. Go back and change the moves. Rewrite the bot's opening message from "You step into the tavern" to "Kael steps into the tavern." Rewrite your last three messages the same way. Send one new message, three sentences, in the perspective you want. Now the copycat sees five examples of the new pattern and zero of the old one — and it copies the only thing on the table.
How to do it
- Pick the perspective you actually want: first person ("I walk in"), second person ("You walk in"), or third person ("Kael walks in"). Third person is Janitor's own recommendation because it best stops the bot from writing your character's actions for you.
- Edit your Persona and rewrite it in that perspective. The persona is fed to the model on every single turn, so a mismatched persona quietly drags the chat back.
- Open the bot's greeting/first message in the chat (the edit or pencil icon) and rewrite it in the target perspective. This is the single highest-leverage change.
- Edit your own last two or three messages into the target perspective.
- Delete or regenerate the bot's most recent replies so the wrong-perspective examples leave the visible history entirely.
- Write your next message in the target perspective and make it two to four sentences long — enough for the model to read the pattern clearly.
- If you want an explicit nudge, add a positive framing rather than a ban: a <system> line like "Narrate in third person, past tense, following only Kael's actions," instead of "don't use second person."
- If it drifts back after a few turns, hunt for leftover wrong-POV text in the greeting, the persona, the character definition, or any advanced-prompt field. Something is still showing it the old pattern.
Key facts
- Janitor AI's interface has no point-of-view or perspective setting as of 2026-07; POV is controlled only by the text in the chat, persona, and character fields.
- Janitor's official character-creation guide recommends third person for the initial message, stating that "third person often provides the best roleplay results in terms of preventing LLMs from speaking for the user."
- Janitor's official writing-style guide states that "JLLM mirrors you" — tone, pacing, and sentence length in your messages shape the bot's replies.
- The first message is a required field and, per Janitor's docs, "sets the tone for the LLM as well, as most will try to follow the pattern of speech and actions dictated in the initial message."
- Janitor's official guidance notes that vague "don't speak for me" instructions often backfire, and that reframing the AI's role works better than prohibition.
- Message edits are retroactive: the model receives only the current text of the conversation each turn, so editing old messages changes all future replies.
▶ The 60-second explainer (script)
You can't change perspective in Janitor AI with a setting, because there isn't one. You change it by editing the text. Here's why. A language model only does one thing: it looks at the words in front of it and predicts what comes next. Perspective isn't a switch — it's a pattern in the words. Janitor's own guide says it plainly: JLLM mirrors you. So if you type "stop writing in second person" while forty messages of second person sit right there on the screen, the pattern wins and your instruction loses. Fix the text instead. Three edits. One: rewrite the bot's greeting message in the perspective you want — that greeting is the loudest style example in the whole chat, and Janitor's docs say most models follow the pattern it sets. Two: rewrite your persona to match. It gets fed to the model every single turn. Three: edit your last two or three messages, then delete the bot's most recent replies so the old pattern is gone from view. Now write one new message, a few sentences long, in your target perspective. Here's the part people miss: editing is retroactive. The model has no memory of what your chat used to say. It only sees the current text. Change the text, and as far as it knows, your story was always written that way.
What authoritative sources say
People also ask
Is there a POV switch in Janitor AI settings?
No. As of July 2026 there is no perspective setting in the interface. Perspective comes entirely from the wording of the greeting, your persona, and the chat history.
Why does the bot keep writing my character's actions for me?
Usually because the greeting or your messages are in second person, which invites the model to narrate "you." Janitor's own guide recommends third person specifically to prevent this.
Do I have to start a new chat to change perspective?
No. Editing works retroactively, because the model only ever sees the current text of the conversation. Fixing the greeting and your last few messages is usually enough.
Does switching to a proxy model fix perspective problems?
Somewhat — stronger models follow instructions more reliably. But every model still mirrors the pattern in front of it, so a second-person greeting will pull any model back toward second person.
How many turns until the new perspective sticks?
Usually one to three, as long as no wrong-perspective text remains in the greeting, persona, or character definition. If it keeps reverting, something old is still in the context.