Run your own AI · System prompts and custom models

System prompts and custom models

thecontinuitylayer.com/open/learn/custom-models

Checked 6 November 2026. US dollars. No affiliate links.

A system prompt is a standing instruction the model reads before every message: who it is talking to, what it is for, how to answer. On your own machine you write it, nobody else adds to it, and you can save it as a model of its own.

By the end

A model saved under your own name, that answers the way you want every time you open it.

Write a good system prompt

Say who you are and what the work is. Many unhelpful refusals and vague answers disappear once the model has that context.
Say how to answer: length, format, tone, what to skip.
Say what not to do, in plain words. Keep the whole thing under a page; a long prompt eats the model's memory.

Set it in your app

LM Studio: paste it into the system prompt box in the chat settings, and save it as a preset to reuse.
Open WebUI: create a model in the Workspace with your prompt and pick it from the model menu; everyone on your server can use it.
Jan: set the instructions for an assistant in its settings.

Save a custom model in Ollama

Make a text file named Modelfile:

FROM gemma4:12b
PARAMETER num_ctx 16384
PARAMETER temperature 0.6
SYSTEM """
You help me run a small bakery. Answer in short steps.
Use Canadian dollars. If you are not sure, say so.
"""

Build it and run it:

ollama create bakery -f Modelfile
ollama run bakery

3
It now shows up as bakery in ollama list and in every app connected to Ollama.

Format from Ollama's Modelfile reference, read on 6 November 2026.

The two settings that matter

Context (num_ctx): how much the model can hold at once. Ollama's default is only a few thousand tokens (4,096 per its FAQ), so long documents and agents need more. Each step up uses more memory.
Temperature: lower (0.2–0.5) for facts and code, higher (0.8–1.0) for brainstorming and fiction.

← Switch models, keep your context Next: Chat with your own documents →