Tag: ai-workflows

Blog Posts

From Fabric User to Pattern Creator: Building Better AI Workflows

Building a simple Fabric MCP server

MCPs that pipe your request through a second LLM, while you're already using the model you wanted, drive me crazy. You get complexity nobody asked for, a conversation that keeps stalling, and a prompt you can't touch once it's running.

So I built the Pattern MCP Server. It hands you the prompt. That's it. No middleman executing it.

To build it I went through all 215 Fabric patterns. Maybe 15% are excellent, the security ones like analyze_malware especially. Another 15% are D/F, and find_female_life_partner boiling a relationship down to an algorithm is icky.

The middle is where the flaws pile up. Cargo-cult prompting, the "think with 1,419 IQ" kind. Constraints wound too tight, "write exactly 16 words per bullet." Roughly 90% ship without a single example, which is wild given examples teach a model more than any instruction does. And the jittery repetition, "DO NOT COMPLAIN" three times in a row.

I added notes on fixing each one.

Everyone needs their own prompt library eventually. Mine's the prompts I've refined for how I actually work. The Pattern MCP Server gives you direct access to Fabric's collection and your custom patterns without the execution overhead. Mix, match, modify on the fly.