Janitor AI prompts that fit the JLLM context budget

Good prompts are clear, short and specific, and that matters more on Janitor AI than on many apps, because of one number on file: the free JLLM model has a context budget of about 8,000-9,000 tokens per chat.

When people search for prompts, they usually mean one of these:

• A character definition: who the bot is, how it talks, what it wants
• A scene opener: where the story starts and what is happening
• A recap: a short reminder you paste in when the bot loses the thread
• A style note: how long replies should be and what the bot should not do

My take: the character definition is where most people go wrong. It is tempting to write pages of backstory. On a budget of roughly 9k tokens, that backstory eats the room the conversation needs, and the bot starts forgetting sooner. I would rather write 10 sharp lines than 100 vague ones.

The official records do not document any special prompt menus, so this guide sticks to what you control: the words you write. Everything below works with the free tier, which the records describe as unlimited JLLM chats and swipes with millions of bots.

Some readers arrive here while moving over from Character AI. Want a fresh start without your old stories first? My guide on clearing out old Character AI chats covers that side.

The table right after this text, The apps side by side, compares Janitor AI with the other apps on this page. It is worth a look before you invest hours in a character, since the apps differ a lot in what they add on top of text.

So a Janitor AI JLLM prompt that fills a third of the budget leaves the chat with two thirds, and older messages fall out of view sooner.

That lines up with the memory notes on file. Other reviewers report drift and forgotten details on free JLLM within 20-30 messages. I cannot promise a shorter definition fixes that, but a bloated one certainly will not help.

Here is what I would keep in a definition, and in this order:

• Name, role and the one trait that makes the character distinct
• How the character speaks, in one or two lines
• The relationship to you and the starting situation
• Two or three rules, such as reply length or never writing your lines

Anything beyond that, I would move into the story itself. Backstory lands better when the character reveals it in a scene than when it sits in a wall of text at the top.

The frontier models also change the game for Janitor AI LLM prompts. The records do not name those models, so I will not tell you which one to pick. My general advice holds for any of them: say what you want plainly, give an example of the tone, and set a reply length. Bigger models tend to follow clear instructions well, which rewards the same lean style.

My view on whether to pay: if your stories keep falling apart after a few dozen messages, the extra context is the feature that addresses it. If short scenes are your thing, the free tier is plenty. For the full picture of plans, memory and limits, my Janitor AI review goes into each one.

• Character: "Mara runs a bookshop in a rainy harbour town. She is warm, teasing and shy about her own poetry. She replies in two or three short paragraphs and never writes my actions or lines for me."
• Scene opener: "It is closing time, a storm has knocked out the power, and Mara lights a candle and asks if I want to wait it out with tea."
• Recap: "Quick recap before we continue: we met at the bookshop, Mara promised to read me a poem, and we are now walking along the harbour at night."
• Style fix: "Please keep replies shorter, stay in character, and let me decide what my character does."

Why these work: each one does a single job. The character line sets voice and limits. The opener gives the model a place and a small problem to react to. The recap is your memory tool, and on the free tier it may be the most useful prompt you own.

The style fix is the one people forget; if a bot rambles or speaks for you, a plain request usually does more than a rewrite of the whole definition.

My one warning: resist adding adjectives. "Warm, teasing and shy" gives a model something to act on. Ten adjectives in a row give it a muddle, and it will pick a few at random. The same goes for rules: two firm ones are followed, ten are half-remembered. If a trait matters, show it in the opener instead of listing it.

These templates are not tied to Janitor AI. The same structure works in any app where you design your own character, and several apps on this page list custom characters as a feature. OurDream AI, in the box below, is one of them.

If you are shopping around, my guide to Janitor AI alternatives compares what each one adds on top of text, from images to voice calls.

That is a poor trade for a chat, and an even worse one if you pay, since Janitor+ subscriptions generally can't be refunded unless the Terms of Service or applicable law require it.

Reading the policy on file is useful here too. Janitor AI's content policy allows adult content. So a Janitor AI NSFW prompt is not a trick to unlock a hidden level; the level is already a matter of the app's policy, and a well-written character and scene do the work.

If a particular model still declines something, that is the model following its own rules, and the honest move is to respect it or choose a different model or app, not to fight it.

My view is that most "jailbreak" searches are really frustration with a bot that will not stay in character or keeps drifting. That is a prompt quality problem, not a rules problem. A tight definition, a clear scene and a short recap fix far more than any clever wording.

Clever wording that tries to trick a model tends to produce worse writing anyway, because the model spends its attention on the trick rather than the story.

If content level is the real issue, pick the app whose policy fits you instead. Among the apps on this page, SpicyChat and OurDream AI list adult content as allowed, while Candy AI and GirlfriendGPT allow it on paid plans only. Choosing that way beats any workaround, and it keeps your account safe.

One last point in favour of playing it straight. A banned account takes your characters with it. If you have spent evenings building a cast you care about, the risk is not worth a single scene. Keep your own copy of your best definitions, too, so you can move them to a new app if you ever want to. For more questions like this, from settings to billing, my help guides collect them in one place.

What I can offer is a workaround that needs no special setting. Put your standing instructions inside the character definition, near the top, in two or three plain lines: reply length, point of view, and anything the bot must not do. That is effectively your own global rule for that character.

If you want the same rules everywhere, keep them in a note and paste them into each character you build.

If you do find a dedicated setting in the app, keep it short for the same reason as everything else here. Whatever it holds, the model has to read it, and it shares the same context budget as the story.

The fix I would reach for first is the recap prompt from earlier: a few lines on who is who and where things stand. Paste it in when you notice drift, not after twenty messages of confusion. Trimming the definition helps too.

If neither is enough and the stories you enjoy are long, that is the point where Janitor+ and its 5x more context start to make sense.

A note on where this advice comes from. My page on how I check the apps explains that the facts here come from official pages and published reviews, while the writing tips are general advice on how these models read text. I keep the two apart on purpose, so you always know which is which.

SpicyChat's free plan lists 4K memory and replies up to 180 tokens, which rewards an even shorter definition than Janitor AI's. Candy AI and OurDream AI both let you create custom characters, so your character line and scene opener carry straight over.

Content rules travel less well. The same character can land very differently on an app with a stricter policy. If you are coming from Character AI for that reason, my piece on whether Character AI will allow NSFW looks at that side of the question.

What does not carry over is anything that leans on images or voice. Janitor AI's feature list shows no images, no voice calls and no group chat, so prompts there are all about text. On Candy AI, images and voice are part of the package, and your prompt may need a line on how the character looks.

That is a different craft, and a fun one, but it is not what the templates above are built for. For a text-only app, every word should serve the story.

Add two rules at most. Then write your first message as if the story has already started, because the model follows your lead more than any instruction.

When things drift, recap. When replies ramble, ask for shorter ones. When you want more depth, add it in the story, not in the setup. And keep a copy of every definition you are proud of, in your own notes, so a good character never depends on a single app.

Start free, get the basics right, and pay only when the context limit is the thing actually holding your stories back.