Mpromp2 — the AI Prompt GeneratorThemeClassic ModernDarkLightSlate

Skills Library / Local & Open Models (Llama, Mistral, GLM)

System Prompts for Local Models

Running Llama, Mistral, or GLM locally (Ollama, LM Studio, llama.cpp) gives you one control the hosted models hide: the raw system prompt and the sampling settings. This page explains how to write system prompts that hold, and gives you a Modelfile-style setup to copy.

How it works

Local serving stacks take a system prompt plus parameters (temperature, top-p, context length). Smaller open models follow shorter, more literal system prompts than frontier models: one role sentence, a numbered rule list, an explicit output contract, and few-shot examples when format matters - examples do more work than adjectives at this scale. In Ollama this is packaged as a Modelfile: FROM the base model, a SYSTEM block, and PARAMETER lines, giving you a named, reusable specialist - the local equivalent of a skill.

Setup, step by step

1. Start from the base model unmodified; change one thing at a time and test
2. Write the SYSTEM block: role in one sentence, rules numbered, output contract last
3. Set temperature deliberately: low (0.2-0.4) for extraction and code, higher only for drafting
4. Add two short examples if the output format matters - they anchor format better than description
5. Save as a named model (Modelfile) so the specialist is one command away

Copy-ready starter file

Field notes

The output contract is the part local models need most: hosted frontier models infer format generously; a local 8B model does exactly what the contract says - so write the contract like one.

Keep going

Mpromp2 — the AI Prompt Generator. Every guide and skill file here is free to use, including commercially.