Prompting Portal: AI that is transparent, collaborative and safe?

A one-hour deep dive at the kulturBdigital conference 2026 (ufaFabrik Berlin, 15 October 2026), together with Christoph Wetzel, ZKM’s digital manager. Earlier the same afternoon the two of us gave the conference closing on the main stage, KI ohne schlechtes Gewissen: Wie es gehen kann (AI without a guilty conscience). The deep dive was the technical half: what is behind it.

What the talk is about

When staff use AI tools unregulated and untrained, the institution carries the risk. It also loses what people learn, because carefully worked-out prompts disappear into proprietary black boxes. ZKM’s answer is ask.zkm.ai, an in-house prompting portal on European infrastructure. The talk tells how it got there, step by step:

• Phase 0: everyone on ChatGPT. In-house knowledge sitting in someone else’s system, good prompts lost.
• Phase 1: our own interface. Open WebUI, with every provider key wired straight into it, and every change made twice.
• Phase 2: one door to the outside. A central LiteLLM gateway holds providers, credentials, fallbacks, rate limits, cost per person and a public CO₂ page.
• Phase 3: in-house knowledge over MCP. Phone book, wikis, website, library catalogue, meeting minutes: the source stays the source, a crawler and a vector database sit in between, and one MCP server per source talks to the model.
• Phase 4: models come and go. Users pick a task instead of a model. Tiles like “general tasks” or “library search” point to an alias, and the alias points to whichever model works best right now. When an academic provider retires a model, nobody notices.

Around that: single sign-on with automatic onboarding, so identity flows through to cost and to the “who am I?” a model can ask; and a three-tier AI driving licence (user, expert, guru), so the people who build good tiles and prompts make them productive for everyone.

Live demo

The middle twenty minutes were a live demo on a conference instance with the same building blocks: chat with tiles, a question to the library catalogue, a document upload, web search with sources, image generation and swapping the model behind a tile live on the gateway.

Rebuild it

The architecture is published as the ask-zkm-ai-blueprint (MIT): gateway, chat interface, MCP servers, proxies and cost/CO₂ tracking, one Docker Compose stack each. There is no one-click install. The repo holds patterns and notes, including what went wrong on the way: tool calls that break depending on the provider, a web search that was broken for months without anyone noticing, an assistant that invented biographies until it was forced to use its tools first.

For the earlier, less technical version of the same story, see the MAI-Tagung talk Prompting ohne schlechtes Gewissen (May 2026).