You are in Janitor AI. You’ve got character cards from your favorite creators, API access that are cheap and flexible via a key, and yet AI responses are still off. Why?

If you have ever pasted a highly upvoted custom prompt into Janitor AI and watched your roleplay get worse, you already know the dirty secret of prompt sharing: there is no such thing as a universal prompt. Janitor AI is essentially a frontend; the quality of your chats depends almost entirely on the model running behind it. A prompt can work well for GLM while causing a mess with DeepSeek, and vice versa.

This motivates me to break down how Janitor AI prompts really work, give you starter prompts for each of the popular proxy models, and explain how to troubleshoot the AI response quality issue so you can have a smooth AI roleplay experience.

How Janitor AI Prompts Work

Before copying anything, it helps to understand the three layers that shape every response as below.

1. The character card: the bot’s personality, backstory, and example dialogue, written by the bot creator.
2. The custom prompt (also called the advanced prompt or system prompt): your instructions that sit on top of every bot you chat with. This is where you control writing style, pacing, formatting, and rules like “never speak for my character.”
3. The model: the actual AI brain generating the text. This can be Janitor’s free built-in model (JLLM) or an external model connected through a proxy, such as DeepSeek, GLM, or Gemini.

The custom prompt goes into the API settings inside a chat. Open a conversation, tap the menu, choose API Settings, and paste your prompt into the custom prompt field. Save, then refresh the page, as Janitor AI often needs a refresh before new settings take effect.

One important mindset shift: a prompt cannot fix a bad character card, and a great card cannot fully survive a weak model. All three layers matter.

Why the Model Matters More Than the Prompt

Every model has its own quirks, trained-in habits, and preferred settings:

• DeepSeek models are expressive and dramatic, but notorious for speaking or acting for your character and for escalating scenes too quickly.
• GLM models produce rich, thoughtful prose, and can get deliberate and straight forward to the point. Also, it responds differently to temperature settings than DeepSeek does. Community testing generally recommends around 0.8 to 0.9 for GLM versus roughly 0.5 to 0.6 for DeepSeek.
• Reasoning models (like DeepSeek’s R-series) may show visible thinking tags before their reply. In general, thinking models take longer, which means you can wait for 20 seconds before getting a response.
• JLLM, Janitor’s free in-house model, has a smaller effective memory and follows long, complex instruction lists less reliably than large proxy models.

This is why the same prompt produces wildly different results across setups. When a prompt “doesn’t work,” there are three possible culprits: the prompt itself is weak, the model cannot follow that style of instruction, or the prompt was simply written for a different model than the one you are using. Always check which model a shared prompt was tested on before blaming yourself or the prompt author.

Starter Prompt for JLLM (No Proxy)

JLLM does best with short, simple, direct instructions. Long multi-section prompts waste its limited context. Try something like this:

Write in third person, past tense. Stay in character at all times. Never write {{user}}’s dialogue, thoughts, or actions. Keep replies between two and four paragraphs. Advance the scene with new actions or dialogue instead of repeating what already happened. Show emotion through body language and behavior rather than stating feelings directly.

You can keep it lean. Every extra rule you add is a rule JLLM might trade away for something else.

Janitor AI Prompts for DeepSeek

DeepSeek (whether accessed through the official API, OpenRouter, or another proxy) is currently the community favorite for expressive, emotionally rich roleplay. Its two biggest flaws are hijacking your character and rushing pacing. A good DeepSeek prompt attacks both problems head-on:

You are the narrator and {{char}} only. {{user}}’s words, actions, decisions, and inner thoughts belong exclusively to the player — never generate them, assume them, or summarize them. End each reply at a natural point that invites {{user}} to respond.

Writing style: immersive third-person prose, two to five paragraphs. Vary sentence length and rhythm. Ground every scene in sensory detail — sound, texture, light, atmosphere. Avoid purple prose and melodrama.

Pacing: let scenes breathe. Do not skip ahead in time, resolve conflicts instantly, or escalate intimacy or tension faster than {{user}}’s messages invite. {{char}} may disagree with, refuse, or challenge {{user}} when it fits their personality. Flawed, realistic behavior is better than constant agreeableness.

Recommended settings from community testing: temperature around 0.5 to 0.6, generous context (16k to 32k), and leave the advanced generation sliders alone unless you know exactly what they do. If DeepSeek still occasionally speaks for you, do not let it slide. Try to edit or regenerate the message immediately, because one uncorrected instance in the chat history teaches the model the habit is acceptable.

Janitor AI Prompts for GLM

GLM has become a serious alternative to DeepSeek for long-form roleplay, and it shines when your prompt works with its natural thinking process instead of fighting it. GLM handles structured, tool-like instructions unusually well:

Before writing, silently consider: What just changed in the scene? What does {{char}} want right now, and what are they hiding? What detail would make this moment vivid? Then write the reply.

Portray {{char}} as a complete person with contradictions — private fears, selfish impulses, moments of unexpected warmth. Their reactions should follow from their established personality and the full history of this conversation, not generic story logic.

Never control {{user}}. Write three to five paragraphs of polished third-person prose. Weave the environment and side characters into the story so the world feels alive around the two leads. Reincorporate earlier details, promises, and objects from the chat history whenever it enriches the scene.

For GLM, run the temperature hotter; around 0.8 to 0.9 works well according to prompt writers who test extensively across both models. This is one of the clearest examples of a prompt-model mismatch trap: run GLM at DeepSeek’s low temperature and it can feel flat; run DeepSeek at GLM’s high temperature and it can spiral into chaos.

A Universal Fallback Prompt

If you switch models often, or you are testing a proxy for the first time, use a minimal prompt that nearly every model can follow, then customize once you know the model’s personality:

Roleplay as {{char}} in third person. Never write for {{user}}. Two to four paragraphs per reply. Match the tone {{user}} sets. Prioritize character consistency, natural dialogue, and forward momentum over flowery description.

Short universal prompts you understand will consistently beat massive multi-thousand-token “prompt operating systems” you copied but never read, and they leave far more context space for the actual story.

Prompt, Model, or Mismatch? A Quick Diagnostic

When your roleplay quality drops, run through this checklist before rewriting anything:

• The bot speaks for you constantly. Partly a model habit (especially DeepSeek). Strengthen the never-write-for-user rule, and edit out every violation so it does not compound.
• Replies feel repetitive or looping. Often a model or settings issue. Edit the last bot message to break the loop, nudge the temperature, or add a rule against reusing phrases and structures.
• The bot ignores your instructions entirely. Likely a mismatch — your prompt may be too long or too complex for the model, or you forgot to refresh after saving settings.
• Great prose but the character feels wrong. That is usually a character card problem, not a prompt problem. No system prompt can fully rescue a thin card.
• The style changed after you switched proxies. Almost certainly a mismatch. Re-tune the temperature and simplify the prompt for the new model before judging it.

The core takeaway: Janitor AI’s output quality lives and dies by the model behind it. Treat every prompt as model-specific until proven otherwise.

If Tuning Prompts Feels Like a Second Job: Platforms With Their Own Trained Models

Some people genuinely enjoy tinkering with proxies, temperatures, and prompt engineering. Others just want good roleplay without the setup. If you fall into the second group, several platforms train and tune their own roleplay models in-house, which means the prompt-model matching work has largely been done for you:

• PolyBuzz offers a huge character library plus creative extras like image generation and visual storytelling tools, making it easy to jump into a scenario with zero configuration.
• SpicyChat is a roleplay-first platform aimed at adult users, built around custom characters, persona control, and open-ended conversations with fewer content restrictions.
• Emochi focuses on anime-style character chat with models tuned for emotionally expressive, character-driven roleplay.
• PovChat runs its own custom models built specifically for roleplay and interactive storytelling, including text-based RPG adventures alongside character chats, with switchable content modes.

None of these replace the flexibility of Janitor AI with a well-matched proxy, but if your goal is immersion rather than configuration, grab one of them and go for the core part. A platform with a purpose-trained roleplay model can deliver a better experience in the first five minutes than an untuned proxy delivers in five hours.

Frequently Asked Questions

Where do I paste custom prompts in Janitor AI? Open any chat, go to the menu, select API Settings, and paste your prompt into the custom prompt field. Save your settings and refresh the page.

Do Janitor AI prompts work without a proxy? Yes, JLLM reads custom prompts too. Just keep them short and simple, because the free model follows long instruction lists less reliably than large proxy models.

Why did a popular prompt make my chats worse? It was probably written and tested on a different model than yours, or your temperature and context settings do not match what the author used. Check the prompt’s intended model first.

Is a longer prompt always better? No. Extremely long prompts eat the context the model needs to remember your story, and models follow a few clear rules far better than fifty vague ones.

What if I can’t find a working prompt? Strip everything back to the minimal universal prompt above and add one rule at a time until something breaks. If even the basics fail, the problem is usually the model or your settings, not the prompt. Try platforms like PolyBuzz, SpicyChat, PovChat AI, and Emochi with models already trained for roleplay, so no prompt hunting is required to get solid results.

Conclusion

The best Janitor AI prompt is not the longest one on Reddit. It is the one written for the model you are actually running, at the settings that model likes, backed by a character card worth roleplaying with. Start with the minimal prompt for your model above, fix one problem at a time, and if the endless tuning stops being fun, remember that platforms with their own trained roleplay models are always an option.