What to put in the JanitorAI custom prompt (with a sample)

A custom prompt you can paste

You are {{char}} in an ongoing roleplay with {{user}}. Write only {{char}}'s words, actions and thoughts, plus any side characters the scene needs. Never write {{user}}'s dialogue, actions, thoughts or feelings; stop and leave room for {{user}} to act. Write in third person, past tense. Keep replies to two to four paragraphs. Show emotion through action and dialogue rather than naming it. Let choices have consequences: injuries, promises, debts and grudges carry forward. Stay true to {{char}}'s personality even when it is inconvenient. Keep the scene moving with one new detail, action or complication per reply.

That is about 110 words. {{char}} and {{user}} are filled in by JanitorAI with the bot's name and your persona's name, so the same prompt works for every bot.

Why each line is there

• "You are {{char}}" sets the job. Without it, some models drift into narrating like an assistant.
• "Never write {{user}}'s dialogue..." is the most valuable line in the box. A model speaking for you is the top complaint in roleplay, and saying where to stop helps more than saying what not to do.
• Point of view and tense stop the reply style from changing halfway through a story.
• Reply length is a soft target. The hard cap is Max new tokens in your generation settings.
• "Show emotion through action" cuts the "a wave of emotion washed over her" filler.
• Consequences and personality lines give the story weight, so characters do not forgive everything in one message.
• "One new detail per reply" keeps scenes from stalling without letting the model rush ahead.

Add-on lines for different bots

Swap or add one line at a time rather than stacking all of them: - Narrator or multi-character bots: "Voice every character in the scene except {{user}}, and give each one a distinct way of speaking." - Slow burn: "Romance builds slowly; no confessions or kisses unless {{user}} clearly moves the story there." - Short, punchy chats: "Keep replies under 120 words, mostly dialogue." - Second person fans: replace the point-of-view line with "Write in second person, addressing {{user}} as you." - Dark or horror stories: "Keep tension high; danger is real and help does not always arrive."

What does not belong in the box

• The bot's personality or backstory. That lives in the bot's own fields and is sent anyway; repeating it wastes tokens and can contradict a bot that was updated.
• Facts about one particular story, such as who knows a secret or where you are. Those go in chat memory, which is per chat; see JanitorAI chat memory.
• Long walls of rules. Every word is sent with every message, and a 1,000-word rulebook buries the lines that matter.
• Rules that fight each other, like "be concise" next to "write long, detailed replies".
• Instructions about sliders. Temperature and length limits are settings, covered in JanitorAI generation settings.

Test a change in five minutes

Paste the prompt, save, reload the page, then go to a chat and swipe the last reply three times. If the bot still writes your actions, move that rule to the start of the prompt and make it more concrete: "End your reply at the moment {{user}} would speak." If replies run long, check Max new tokens before adding more words to the prompt.

Change one line per test. When several lines change at once you can't tell which one helped. With gemma-4-12b or gemma-4-31b on UnsavedAI, plain direct sentences like the sample above tend to work better than clever formatting. For the general idea behind this box, see what a system prompt is.

Questions people ask