Good Janitor AI prompts are short, specific and written as instructions the model can actually follow. Most bad chats come from the same few causes: a bloated character definition, a first message that narrates your character, and “don’t do X” rules that backfire. This guide covers every place a prompt goes on Janitor, with copy-paste examples for each, plus fixes for the bot speaking for you, repetition and bad pacing.
Cherrypop.ai publishes this guide, and we make a competing AI companion app. Based on each vendor’s pricing and help pages and published user reports, as of September 2026. The example characters are original, so reuse them freely.
Where Janitor AI prompts go (and why it matters)
Some fields are “permanent tokens” the model sees every turn; the chat itself is temporary, and the oldest messages drop out as it grows.
| Field | What it’s for | Seen every turn? |
|---|---|---|
| Personality (required) | Traits, appearance, backstory, speech. Janitor suggests putting dialogue examples here too. | Yes |
| Scenario (optional) | Setting, time period, a short line on how the character knows your persona. | Yes |
| Initial Message (required) | The opening message. Sets the tone, length and point of view the model copies. | Only while it’s still in context |
| Persona | Who you are in the story. | Yes |
| Chat Memory | Running notes: where you are, what happened, relationship status. | Yes |
| Advanced / Custom Prompt | Your system prompt: rules for style, length and roles. | Yes |
| OOC or task lines in chat | One-off instructions mid-roleplay. | Until they scroll out |
The budget is tight. Janitor’s help center puts the free JLLM model’s context at roughly 8,000 to 9,000 tokens (about 1,000 tokens per 750 words). Its token guide says most well-made bots work best with under 1,500 tokens of permanent info and that past 2,000 you’re “skating on thin ice”; the character creation page recommends staying under 2,500. Every setup token is one less for your chat.
Three rules that make every prompt work better
- Say what to do, not what to avoid. Janitor’s Advanced Prompting 101 guide warns that “No blood” is still “blood”: negative words hand the model the very idea you’re trying to keep out. Flip the rule: instead of “don’t be cheerful”, write “{{char}} is dry and deadpan.”
- Use firm verbs. “Feel free to describe the setting” reads as optional. “Describe the setting in vivid detail” doesn’t. Use, maintain, describe, write.
- Say it once. The same guide calls repetition “noise”. Ranges also work better than limits: “limit to 1 to 3 paragraphs” beats “no more than three paragraphs”.
Character definition prompts (the Personality box)
Write concrete traits, not adjectives. “Nice” gives the model nothing; “offers you her jacket, then pretends she was too warm anyway” gives it a behavior to repeat. Here are three formats, from easiest to most compact.
1. Plain-prose definition
Vera Kaine is a 31-year-old salvage pilot who flies a battered cargo ship called the Magpie. She is sarcastic, fiercely competent and allergic to being thanked. She shows affection through actions: fixing things, sharing food, standing between {{user}} and trouble. She speaks in short, clipped sentences and swears when nervous. Appearance: tall, cropped black hair, grease on her hands, a scar through her left eyebrow. She owes money to the wrong people and hides it badly.2. Bracket template (compact)
[Theo Marsh;
Age= 35
Job= owns a late-night bar on the harbor
Appearance= broad shoulders, rolled sleeves, grey at the temples, warm brown eyes
Personality= patient, observant, teasing, protective, slow to open up
Speech= low voice, dry jokes, calls regulars "trouble"
Likes= jazz records, quiet closing hours, people who tip in stories
Dislikes= liars, loud phones
Secret= is quietly selling the bar and hasn't told anyone]3. Dialogue examples (add under the definition)
Example lines teach the voice faster than any description. Here are four for Theo, across a few moods.
Dialogue examples (for tone; write fresh lines in the same voice):
Greeting: "You again. Sit. You look like you need something strong and a bad idea."
Annoyed: "I said I'm fine. Hand me that crate and quit looking at me like that."
Flirting: "Careful. Keep smiling at me like that and I'll start charging you rent."
Vulnerable: "Everyone leaves the harbor eventually. I just figured I'd beat them to it."Scenario prompts: 5 setups you can copy
Keep the Scenario box to what’s always true: setting, era, and the relationship. Changing plot details belong in Chat Memory.
4. Rivals to lovers
{{char}} and {{user}} are rival architects, both 30s, competing for the same museum commission in Lisbon. They have history from a conference two years ago that neither mentions.5. Fantasy tavern
A frontier town in a low-magic kingdom. {{char}} runs the only inn and knows every secret in town. {{user}} arrives at night, soaked, with a sealed letter addressed to {{char}}.6. Second chance
Present-day Chicago, winter. {{char}} and {{user}} were engaged eight years ago. They meet again as best man and maid of honor at a mutual friend's wedding.7. Sci-fi crew
A deep-space freighter, six weeks from port. {{char}} is the ship's captain; {{user}} is the new engineer hired after the last one vanished.8. Slow-burn contract
{{char}}, a 38-year-old novelist, hires {{user}} as a live-in assistant for one summer on a remote coast. Strictly professional, at first.First message prompts that stop problems before they start
The first message sets the length, style and point of view the model copies for the rest of the chat. Janitor’s creation guide says third person “often provides the best roleplay results” for keeping the model from speaking for you. A good opener places the character, shows one trait in action, and ends on something you can answer. It never writes your character’s actions or feelings. Prompt 9 is a Theo opener.
Rain hammers the harbor windows. Theo is stacking chairs when the door opens, and he doesn't look up. "Kitchen's closed," he says, then catches a glimpse of who's dripping on his floor. He sets the chair down slowly. "Well. Didn't think I'd see you back here." He nods at the stool nearest the bar, already reaching for two glasses. "You going to tell me what happened, or do I have to guess?"No “you feel nervous”, no “you sit down”: every choice is yours. A five-paragraph intro teaches five-paragraph replies.
Persona prompts: tell the bot who you are
Janitor’s persona guide suggests around 500 tokens or less, wrapped in tags so the model reads it as facts about you, not instructions. Stick to stable facts: name, look, role, how others see you. A long list of your own personality traits invites the model to act them out for you. Prompt 10 is Mara, Theo’s old flame.
<user_persona>Mara Ellis:
- 29, woman
- short red hair, freckles
- old army jacket, battered film camera on a strap
- freelance photographer, back in the harbor town after eight years away
- known locally as the one who left</user_persona>Advanced prompt examples (Janitor’s system prompts)
Janitor calls system prompts “Advanced Prompts”, and its token guide counts them as permanent tokens on JLLM too. On a proxy, paste yours into the Custom Prompt field of your API Settings configuration. Combine one or two; a stack of ten rules is how prompts break.
11. Role and point of view
You are the narrator of an interactive story. You write {{char}} and any side characters: their dialogue, actions, thoughts and the world around them. {{user}} is played by the human, who writes all of {{user}}'s words, actions and feelings. Write in third person, present tense, matching the opening message. End each reply on {{char}}'s action, line or question, leaving {{user}}'s response open.12. Length and format
Limit each reply to 1 to 3 paragraphs. Put dialogue in "quotes" and actions in *asterisks*. Focus on one beat per reply: a reaction, a line, or a move.13. Slow-burn pacing
This is a slow-burn romance. Build attraction through small moments: glances, near-touches, banter, things left unsaid. {{char}} keeps their guard up and lets trust grow over many scenes. Advance the relationship one small step at a time, at the pace {{user}} sets.14. Vivid prose
Write grounded, sensory prose: sounds, textures, temperature, light. Keep metaphors rare and fresh. Vary sentence length and openings. Move the scene forward in every reply with a new detail, choice or complication.15. Mature scenes, consent built in
Both characters are consenting adults. When the story turns intimate, write it with frank, direct language and physical detail, building gradually. {{char}} reads {{user}}'s cues, checks in naturally in character, and follows {{user}}'s lead on pace and limits.Putting it together: a complete advanced prompt
Here are prompts 11, 12 and 14 merged into one block you can paste as-is. Add 13 or 15 if the story calls for it.
You are the narrator of an interactive story. You write {{char}}, side characters and the world: dialogue, actions, thoughts, setting. {{user}} is played by the human, who writes all of {{user}}'s words, actions and feelings. Write in third person, present tense. Limit each reply to 1 to 3 paragraphs, dialogue in "quotes", actions in *asterisks*. Write grounded, sensory prose and move the scene forward in every reply with a new detail, choice or complication. End each reply on {{char}}'s action, line or question.Whether a bot allows explicit content depends on its Limited or Limitless tag and on Janitor’s content guidelines, so check both before you build a mature bot.
Chat Memory template (with a filled-in example)
Janitor’s help center recommends a structured memory with six headings, short bullets and one verb tense. List facts, not scenes, and describe what the character is, not what they always do. Prompt 16 fills it in for Theo and Mara.
Environment:
- the bar after closing, storm outside
Relationship Dynamic:
- {{user}} and {{char}} are old flames with unresolved feelings
Current Plot Points:
- {{char}} is secretly selling the bar
{{char}} notes:
- inventory: {{user}}'s old photo, hidden behind the register
{{user}} notes:
- outfit: old army jacket, film camera
Important Past Events:
- {{user}} left town eight years ago without saying goodbyeFixing common problems with Janitor AI prompts
The bot speaks or acts for you
It’s one of the most common complaints, and Janitor’s own guide says most “do not speak for user” lines don’t work: the model reads them as putting your character in the spotlight and writes about you more. What helps, per the same guide, is giving the AI a clear role (prompt 11 above) and writing fuller replies yourself. If you only send a line of dialogue, the model fills the gap by writing your actions.
- Use a third-person first message that doesn’t narrate you.
- Write what your character says, does and notices in each turn.
- Edit out any line where the bot wrote for you, or reroll. User reports say the pattern sticks if you let it slide once.
- If the chat is badly off track, start a new one. Janitor’s guide says clean slates “solve more problems than you’d expect”.
Repetition and looping
Repetition usually means the context is full and the model is copying the dominant pattern. Janitor’s token guide lists “glitchy, repetitive responses” as a symptom of overloaded memory. Trim your permanent fields, delete the repeated replies so they stop feeding the loop, and add a fresh recap to Chat Memory. For a one-off nudge, use prompt 17:
(OOC: Vary sentence structure and openings. Move the scene forward with a new action or detail.)Temperature matters too. Janitor’s temperature guide suggests starting at 0.6, raising it by 0.1 if replies feel bland, and staying below 1.0.
Pacing is too fast
If the story jumps three scenes ahead in one reply, add the slow-burn prompt (13) and the one-beat format rule (12). Mid-chat, try prompt 18:
(OOC: Slow down. Stay in this moment and write only {{char}}'s next reaction.)Replies are too long
In the model settings tab, set Response Length to Short. That switches to JLLM-short, which Janitor says gives replies about 70% shorter, and it’s open to all users. The model also mirrors you, so shorter messages get shorter replies. User reports say a low max-token cap cuts replies mid-sentence instead, so steer length with the prompt.
The bot forgets things
Once something scrolls out of context, the model can’t see it unless you reintroduce it. Janitor’s memory guide suggests asking for a summary every so often and pasting it into Chat Memory. It recommends this out-of-character format, wrapped in system tags for extra weight. Prompt 19 lists events; prompt 20 tracks the relationship.
<system>task: pause chat|roleplay, answer query(outside of roleplay, succinct summary(list of all impactful events in roleplay(for each event listed: +their effects)))</system><system>task: pause chat|roleplay, answer query(outside of roleplay, succinct summary(include relationship changes between {{user}} and {{char}}))</system>The “pause chat|roleplay” part is what stops the bot answering in character. User reports suggest deleting these messages once you have the summary, so the chat doesn’t drift into bullet points or meta commentary. If you’re wondering what Janitor does with your chats, see is Janitor AI safe.
JLLM vs a proxy: does your prompt need to change?
JanitorLLM (JLLM) is Janitor’s own model and the free default. You can also connect other models through a proxy in API Settings, or, for janitor+ subscribers on the web, through janitor+ Router, a newer option that bills per token from an in-app wallet. For subscribers, Janitor still recommends its JLLM+ model over Router for everyday roleplay.
The same prompt can behave differently on different models. Janitor’s prompting guide gives the example that the word “emphasis” does very different things in DeepSeek and JLLM. For proxies, its beginner proxy guide has recommended that most models work best with a context size of 16,384 tokens, and noted that going higher can make them slower and more forgetful. Test on the model you’ll actually use. If you’re weighing the free and paid options, see whether Janitor AI is really free.
Want ready-made characters instead?
Rather skip token budgets and proxy settings? Cherrypop.ai is an uncensored AI chat and roleplay platform with a large catalog of ready-made characters (female, male and trans, in realistic and anime styles) across many personality types, from tsundere to kuudere. If none fits, the create-your-own wizard walks you through building one.
It also generates images of your character, including NSFW. There’s a free tier, and paid plans unlock more messages, images and features (see pricing on the site). It runs in the browser, including on mobile, with no app to install, and it’s 18+ only, with age checks where the law requires them (go.cam). For more options, see our list of Janitor AI alternatives.
FAQ
How long should a Janitor AI character definition be?
Janitor’s help center says most bots work best with under 1,500 tokens of permanent info and recommends staying under 2,500. That total includes personality, scenario, chat memory, persona and any advanced prompt, not just the definition.