The answer first: there is no POV button
A perspective change in Janitor AI is done with prompts, because the platform has no POV feature: nothing in the help center offers a first-person or third-person setting, and the only official article that touches the topic uses a POV instruction as an example of phrasing, not as a feature. So the real question is not what to type; it is where to put it. Janitor gives you three placements with three different lifespans, and the whole technique fits on the card above.
What people mean by “perspective change” is usually one of two problems, and both use the same placements: making the bot narrate in a different person (first to third, or writing from the character’s own point of view), and the more common complaint, the bot speaking and acting as you. The prompts differ by a sentence; the mechanics are identical.
Placement 1: the persona or advanced prompt (permanent)
The persistent field is the anchor, and it outranks any mid-chat correction for durability. At character or persona setup, in the advanced prompt:
Always narrate in third person. Never write dialogue, thoughts, or actions for {{user}}.Swap the first sentence for whatever perspective you actually want (“Narrate from {{char}}‘s point of view, in first person” works the same way). One phrasing note from the official prompting guide: bare negatives are weak on their own, so pair the “never” with the positive instruction, as above, rather than shipping the prohibition alone.
Placement 2: response prefill (per chat)
Generation Rules, then Response Prefill, seeds the shape of every reply in that chat:
[{{char}} replies in third person only, from {{char}}'s point of view.]It works from the first message, and it carries one documented landmine: DeepSeek V3.2 rejects prefilled text and can output gibberish while a prefill is set. If you run a DeepSeek proxy and your replies degrade, clear this field before you debug anything else.
Placement 3: the OOC line (instant, temporary)
Mid-scene drift gets fixed inline, appended to your own reply. Of the three community OOC syntaxes, the system-tag form ranks highest:
<system>From now on, write only {{char}}'s actions and dialogue, in third person.</system>It obeys immediately and it fades, because it lives in the context window and scrolls out with it. That is not a bug in your prompt; it is what the placement is for. Use it to reset the scene, and rely on placement 1 to keep it reset. If the bot relapses, one more system line plus one edited reply (rewrite a single response into the perspective you want; the model copies the visible pattern) beats sending the same correction five times.
When it is the model, not the prompt
Perspective discipline varies by brain. The free built-in model loses instructions faster than the big proxied ones; if your third-person rule will not hold no matter where you place it, the fix might be a better model behind the proxy rather than a better sentence. And if the whole site is ignoring everyone today, that is a different page: is Janitor down.
Honest closing note: some apps make this entire page unnecessary, because perspective is a setting, not a negotiation. On Candy AI, the strongest texting of my test cycle, the model held character and perspective without any of the scaffolding above; my JanitorAI alternatives page maps which platforms still need the scaffolding and which do not.
Sources
- JanitorAI help center: Advanced Prompting 101, the official prompting guidance, accessed Aug 2, 2026
- Community technique documentation, Jun 29 2026 (updated Jul 14 2026) and Mar 2026 guides mapping the three placements and OOC syntaxes, as recorded in the research archive
- The three prompt placements verified against JanitorAI’s public field names (Persona, Advanced Prompt, Generation Rules, Response Prefill), Aug 2, 2026