Michael Limberger

A blank brain with a job description

This is where OpenWebUI goes from a nice ChatGPT clone to something more useful. Workspace lets you create custom model configurations that act like specialized agents. Base models from Ollama are blank slates. A custom model is that same brain with a job, a personality, and settings tuned for a task.

Click Workspace in the sidebar. You will see three tabs: Models (custom configurations), Knowledge (document collections for RAG, where the model answers from your files), and Prompts (reusable templates). We are on Models.

Click + or Create a Model. Name is what shows in the dropdown: Code Reviewer, Writing Assistant, Explain Like I'm 5. Model ID is auto-generated; you can shorten it. Description is a note to future you. Avatar is cosmetic. GIFs are allowed if you are fancy.

Base Model is the brain. The custom model does not create a new AI. It wraps an existing one. Code Reviewer might use qwen3:30b-a3b with a prompt about quality and security. For general chat, qwen3:30b-a3b is a solid pick. For long documents, try qwen3-next if you have 64GB. For images, use qwen3-vl or mistral-small3.2. For fast and light, qwen3-vl:4b works.

The silent first message

The system prompt is the most important field. It defines who this model is, and it loads silently before the user says anything.

You are a code reviewer. When the user shares code, analyze it for:
bugs, security issues, performance problems, and readability.
Be direct and specific. Point to exact lines. Suggest fixes with
code examples. If the code looks good, say so briefly.

Dynamic variables use Jinja2 syntax: {{CURRENT_DATE}}, {{CURRENT_TIME}}, {{USER_NAME}}. Example: "Today is {{CURRENT_DATE}}. The user's name is {{USER_NAME}}. Greet them by name when appropriate."

Prompt suggestions are clickable chips on a new conversation. For Code Reviewer you might offer "Review this Python function for security issues." For Explain Like I'm 5 you might offer "Why is the sky blue?" They are not required. They just make the custom model feel finished.

Capability toggles include Vision (needs a vision base model), Web Search, Image Generation (needs DALL-E or similar), and File Upload. A code reviewer probably does not need image generation. A research assistant probably wants web search. Knowledge bindings attach document collections so the model has them every conversation: FAQ bot, study assistant, coding helper with project docs. Tools attach calculators, code runners, and external systems, both built-in and custom. The option is there when you need it.

Save, and the custom model appears in the dropdown. Same base model, different prompts, as many as you want. One for code, one for writing, one for your grandmother.

Home Chef, as a walk-through

Name it Home Chef. Base model: qwen3:30b-a3b. System prompt: an experienced home cook focused on practical everyday cooking, always include prep and cook time, assume a standard kitchen, respect dietary restrictions going forward, and keep it conversational because nobody wants an essay on boiling an egg. Starters: chicken thighs and rice, a quick weeknight dinner for two, pasta sauce from scratch, a substitute for heavy cream. Select Home Chef and talk about dinner. The prompt shapes the replies without you explaining the rules every time. Next: what makes a good system prompt.