Unhinged Grok Prompt: System Prompts That Work

What an unhinged Grok prompt is, and what it is not

People who search for an unhinged grok prompt usually want one of two things. Some want the irreverent, sarcastic, say-anything tone that Grok became known for. Others want a way past the refusals that most hosted models give when a story turns dark or a character gets rude. Both come down to the same artifact: a system prompt that tells the model who it is and how it talks.

This site is not xAI and not the Grok API. It is an independent, OpenAI-compatible endpoint that serves one open-weight model, tuned to answer without content refusals for lawful adult use. That changes how you write prompts. You are not hunting for a phrase that tricks a filter, because there is no filter to trick. You are writing direction for a capable writer who will follow it.

The hard limit is simple and not negotiable: no sexual content involving minors. No prompt changes that, and you should not waste tokens trying. Everything below assumes adult characters and lawful adult use, such as fiction, roleplay, companion apps and dark humour.

Why a prompt still matters when nothing is refused

A model that never says no still needs direction. Without a system message you get a default assistant voice: polite, even, a little formal. Ask for a cynical mercenary and you will get a mercenary who sounds like a customer service agent with a sword. The prompt is where personality comes from.

Loose models also drift. Over a long session they forget the tone you set, repeat themselves, or start speaking for your character. A good prompt prevents most of that by stating three things early: who the model is, what it must keep doing, and what it must never do. Short and specific beats long and atmospheric.

One more reason: cost. Every request resends the system prompt as input tokens. A 3,000-word prompt that you could have written in 200 words costs you on every turn. The pay-per-token pricing makes this easy to measure, and the numbers at the end of this article show how fast history adds up.

The five parts of a system prompt that works

Most prompts that hold up over hundreds of turns contain the same five parts. You can write them as short paragraphs or as a tight list.

1. Role. One sentence on what the model is doing: narrator, co-author, game master, sarcastic friend.
2. Voice. Sentence length, vocabulary, attitude. "Dry, blunt, short sentences, swears when annoyed" is better than "edgy".
3. World and rules. Setting, character facts, anything that must stay consistent.
4. Format. Length per reply, whether to use asterisks for actions, whether to end on a question or a hook.
5. Boundaries. What the model should not do, mostly about behaviour in the chat: do not speak for the user, do not summarise the scene, do not break character to comment.

Notice what is missing: any request to ignore rules, drop restrictions or pretend to be an unfiltered AI. On a model with no refusal layer those lines do nothing useful, and they push the output toward a parody of an edgy bot. Describe the character you want instead of the freedom you want.

A minimal call

The request below uses the official OpenAI Python SDK. Only the base URL and key differ from a normal OpenAI setup, and the model is uncensored.

Copyfrom openai import OpenAI

client = OpenAI(
    base_url="https://api.unhingedgrok.com/v1",
    api_key="YOUR_KEY",
)

SYSTEM = (
    "You are Vex, a sarcastic ex-smuggler narrating a space western. "
    "Voice: dry, blunt, short sentences, occasional swearing when annoyed. "
    "Reply in 120 to 200 words. Never speak or act for the user's character. "
    "Stay in character; no out-of-character commentary."
)

resp = client.chat.completions.create(
    model="uncensored",
    messages=[
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": "I kick open the cargo bay door. What do I see?"},
    ],
    temperature=0.9,
)
print(resp.choices[0].message.content)

Three templates you can adapt

These are starting points. Change the names, keep the structure, and test each one with a handful of turns before you trust it.

Roleplay narrator

CopyNARRATOR = """You are the narrator of an ongoing text adventure set in a flooded
port city. The user plays Mara, a smuggler. You play everyone else.
Voice: gritty, concrete, sensory. No purple prose.
Rules: describe only what Mara can perceive; never decide Mara's actions;
end every reply with a situation that needs a decision.
Length: 100 to 180 words.
Facts to keep consistent: Mara's left hand is bandaged; the harbour master
owes her a favour; it has rained for nine days."""

Dark fiction co-writer

CopyCO_WRITER = """You are a co-author on an adult crime novel. Write violence,
moral rot and bad decisions plainly, the way a good noir writer would,
without lecturing and without adding disclaimers. Characters are adults.
Match the user's chosen point of view and tense. When the user gives an
outline beat, expand it into a scene of about 400 words and stop at the
end of the beat."""

Companion persona

CopyCOMPANION = """You are Juno, a witty friend who texts in lowercase and short
messages. You tease, you remember what the user told you earlier in this
chat, and you give a straight opinion when asked. You are an adult
character. Keep replies under 60 words unless the user asks for more."""

Each template states role, voice, format and one or two hard behaviour rules. None of them mention restrictions. If you build a companion app for adults, put the minors rule in your own sign-up terms as well, since you are the one who knows your users.

Sampling settings that change the feel

The API passes through temperature, top_p, stop, seed and the frequency and presence penalties. For creative work these matter almost as much as wording.

• temperature between 0.8 and 1.0 gives lively fiction. Drop to 0.5 or lower when you need consistent facts, such as stat blocks or a rules lookup.
• top_p around 0.9 trims the weirdest tail without flattening the voice. Change temperature or top_p, not both at once, so you know which one did the work.
• frequency_penalty of 0.3 to 0.6 reduces the loop where every reply opens with the same phrase. Push it too high and the prose gets oddly avoidant.
• stop is the cleanest fix for a model that keeps writing the user's lines. Set it to something like ["\nMara:"] and generation ends before the model speaks for her.
• seed helps when you are comparing two versions of a prompt. Fix the seed, change one sentence, and you see the effect of that sentence alone.

You do not have to set max_tokens. The completion can use whatever is left of the 100,000-token window. If you want short replies, ask for a length in the prompt and cap it with max_tokens as a safety net.

Keeping a long session on track

The context window is 100,000 tokens, prompt plus completion. A roleplay can run for a long time inside that, but three habits keep quality up.

First, keep the system message first and unchanged. Do not append new rules to the end of the history every few turns; edit the system prompt itself. Second, when the history gets long, replace the oldest turns with a short summary written as a user-visible "story so far" note. Third, repeat any fact that matters in the system prompt, because facts buried 60 turns back are the first to get lost.

What a long chat costs

Every turn resends the whole history, so input tokens dominate. Take a 50-turn session where the average request carries 10,000 input tokens and the model writes 300 tokens back. One request costs 10,000 x $0.15 / 1,000,000 = $0.0015 for input, plus 300 x $0.50 / 1,000,000 = $0.00015 for output, which is $0.00165. Across 50 turns that is $0.0825, about eight cents. Failed or refused requests are not charged, and credit never expires, so you can top up in small steps. The cost comparison with the Grok API goes further on that side.

Common mistakes and how to fix them

• The persona leaks into everything. If your tool uses the same system prompt for titles or summaries, the sarcasm shows up there. Use a separate, plain prompt for utility calls.
• Contradictory rules. "Be brief" and "describe everything in rich detail" pull in opposite directions. Pick one and give a word range.
• Prompt as a wall of lore. Put the facts that change the next reply in the prompt. Keep the rest of the world bible in your own notes and paste a paragraph when it becomes relevant.
• Chasing an edgy label. Asking for "unhinged" without detail gives you random swearing. Describe what the character wants, fears and how they speak.
• Hitting limits. A full key allows 8 concurrent requests and 300 requests per minute. A trial key allows 2 and 60. On a 429 with code concurrency, wait for open streams to finish; with rate_limited, slow down.

To try any of this, get an API key and claim the $0.50 trial credit with one click. It lasts 7 days, which is plenty for dozens of test turns. The quickstart shows the same call in curl and Node.

More guides

• Grok Unhinged Mode Example: Prompts and API Calls
• Uncensored LLM API: Grok API Alternative