How to Stop Janitor AI Talking for You (and Fix It)

Short answer: to stop a Janitor AI bot from speaking or acting for you, do not rely on a "don't talk for me" line. Janitor AI's official guide says that instruction often backfires. Instead, define the AI as a narrator, write your own character's actions in enough detail that nothing is left to fill, and keep your writing style consistent. If it still happens, start a new chat.

The same tools control the other half of the problem: making a bot reply in the voice, point of view and length you want. This guide covers both, with prompts you can copy, a test you can run, and notes for people who connect their own model such as DeepSeek. Janitor AI's guide was last updated on August 11, 2026.

Why "don't talk for me" fails

The official writing guide explains the mechanism. When you tell the AI "don't talk for {{user}}", it reads the line literally, as if you were pausing to let your character take centre stage. The only way the AI knows to let your character drive the story is to write about them, so the command often makes it write more about your character, not less.

The guide's conclusion: "Instead of trying to keep the AI quiet by saying what not to do, it's more reliable to clearly define what it is." It also says that wrapping the same instruction in a system tag usually does not help. In plain terms, a rule that names the unwanted behaviour places that behaviour in the model's context, and the model continues patterns.

Fix 1: give the AI a narrator identity

Assign the AI a job that is separate from you and from the characters. The guide notes that the AI does not only act out characters, it also creates an invisible narrator behind the scenes. A clear narrator identity gives it a lane to stay in. A starting point:

You are the narrator and the non-player characters in this story.
You do not control {{user}}. Describe the world and other characters only.
End each reply where {{user}} can act.

The first line defines what the AI is. The second sets a boundary in plain words. Keep it short, because a long rules block gets ignored. Put the role first, before any style rules.

Fix 2: leave no gap for the bot to fill

The guide says that if you write only dialogue, with no actions, thoughts or reactions, the AI may jump in to "complete" the story. It is not being pushy. It is trying to keep the narrative alive, and it fills whatever space you leave open.

Compare two messages.

Gap left open	Gap closed
"Fine. Let's go."	"I pick up the lantern and check the door latch before answering. 'Fine,' I say, keeping my voice flat. 'Let's go.' I wait at the threshold for the guide to move first."

The second message shows what your character says, does and intends, and it ends on a clear stopping point. A single paragraph like that is often enough. You do not need to write a novel each turn.

Fix 3: write the way you want replies written

The guide says "JLLM mirrors you", meaning that if you write flowery descriptions the bot will too, and if you write snappy dialogue with short action beats it will follow. That note is about Janitor's own model, but mirroring is common to chat models. Match the length, tense and point of view you want back.

You want	Write like this
Short replies	Short messages, with the action in one or two sentences
Second-person narration	Write your turns in second person
Formal tone	Keep your own tone formal
Detailed scenes	Describe the setting and sensations in your turns

The guide also says consistency matters more than perfection. The more steadily you write, the more steadily the AI responds.

Fix 4: use instruction tags for out-of-character requests

When you need to step outside the story to ask for a summary or a scene change, the guide recommends an instruction format that marks the line as a request and not as story text:

task: pause chat|roleplay, answer query(detailed summary(STUFF HERE))

The pause chat|roleplay part tells the bot to leave character mode, and answer query(...) tells it exactly what you want. The guide gives variants you can reuse:

• Recap events: task: pause chat|roleplay, answer query(detailed summary(relationship between A and B))
• Check lore: task: pause chat|roleplay, answer query(detailed summary(analyze lore to evaluate its internal consistency))
• Change scene: task: pause chat|roleplay, seamlessly switch scene to: [description]
• List the location: task: pause chat|roleplay, answer query(detailed summary(describe current location (format: succinct nested list)))

A summary like this is also a memory tool. Janitor AI's memory guide suggests asking the model for a recap every so often and then saving it as the new chat memory, which keeps a long chat coherent without resending every message.

Test whether a fix works

Do not judge a prompt from one reply. Use ten messages.

1. Write ten turns of your own, each with a clear stopping point, and save them.
2. Run them against the current setup and count how many bot replies include your character's actions, words or thoughts.
3. Change one thing, for example add the narrator role.
4. Run the same ten turns again and recount.

A drop from, say, six violations to one means the change works. If the count does not move, revert it and try the next fix. Changing several things at once tells you nothing about which one helped.

If you use your own model through a proxy

These tips come from guidance written for Janitor AI's built-in model. If you connect DeepSeek or another model, test them rather than assuming they transfer. Three details matter for DeepSeek.

• Temperature may do nothing. DeepSeek's thinking mode guide says thinking mode ignores temperature, and thinking is on by default. Do not count on the slider to fix repetition.
• Length costs money. Set a max-token cap to control reply length. Output tokens cost more than input tokens on every provider, and DeepSeek bills thinking tokens as output.
• A stable prompt is cheaper. Keep the narrator role at the top of the prompt and do not edit it between messages, so DeepSeek's cache keeps covering it. See DeepSeek API pricing: what 1,000 messages really cost.

Our guide to the best DeepSeek model and prompts for Janitor AI covers the settings and a fuller template, and how to set up DeepSeek on Janitor AI covers the connection.

Quick troubleshooting

Problem	Try
Bot writes your actions	Narrator role first, then more detailed user turns
Replies too long	Shorter messages, lower max tokens
Bot ignores the rule	Shorten the prompt, put the role first, drop the "don't" lines
Style drifts	Keep your own writing consistent in tense and length
Chat feels stuck	Ask for a recap with an instruction tag, or start a new chat
Nothing helps	Start a new chat. The guide says clean slates solve more problems than you would expect

Sources

Janitor AI: Writing Style and Talking to the Bot, Janitor AI: Chat Memory and Context Management, Janitor AI: Advanced Prompting 101, DeepSeek: Thinking Mode.