I got sick of LLM pleasantries and disclaimers, so I built a system prompt to fix it (SutniPrompt v0.1.0-alpha)
TL;DR: Tired of LLM fluff and "As an AI..." disclaimers. Built SutniPrompt (v0.1.0-alpha), a system framework that forces Claude, Gemini, and GPT into a strict analytical mode. It kills pleasantries, enforces structural markdown, mandates Wikipedia citations, and features a "Mandatory Halt" that stops hallucinations on vague prompts by forcing the AI to ask clarifying questions.
---
Hey everyone,
Like a lot of you, I was getting incredibly frustrated with how commercial LLMs (GPT, Claude, Gemini) constantly pad their answers with unnecessary pleasantries, safetyism, or those endless "As an AI language model..." disclaimers. I just wanted an analytical tool that gives me straight answers and frameworks, not a chatty assistant.
So, I’ve been working on a structured system instruction framework called SutniPrompt. I just pushed v0.1.0-alpha to GitHub.
Here is what it actually does to the model:
- Kills the fluff: Forces "stealth mode". It executes silently without justifying its tone or faking empathy.
- Forces analytical structure: Mandates clean Markdown and prioritizes mental models over dogmatic, definitive conclusions.
- The "Mandatory Halt": This is my favorite part. If a prompt is too broad or asks for a plan based on non-existent info, the prompt forbids the LLM from hallucinating a massive wall of text. Instead, it forces the model to stop and output ONLY 2-3 clarifying questions.
- Fact-checking mandate: Forces the model to always end the response with exactly one relevant Wikipedia link.
How to use it: It’s a bit heavy, so deployment depends on the UI. It works natively in Claude’s System Prompt settings. For Gemini, I’ve documented a modular copy-paste method. For ChatGPT, it's currently best used as an initialization prompt at the start of a chat (I'm working on a minified version that fits perfectly into GPT's Custom Instructions limit for the next releases).
I’d love for some of you prompt engineers to test it out, try to break the gating logic, and let me know what you think.
I'm already working on next updates, they will come really soon, aiming at a full release. I'll document the progress on Github with multiple pre-releases.
Repo and full documentation here: https://github.com/sutnip/sutniprompt
Cheers!