Building a Personal Prompt Library

I built the Pattern MCP Server partly to learn the ropes of MCP development. Mostly, though, I built it because the existing options, Fabric's MCP included, run your prompt through Fabric's configured LLM and relay the results back to you. It drives me crazy that so many MCPs proxy calls through another LLM when I'm already using the LLM I want. The relay adds complexity and breaks conversation flow (and context!). It also means you can't tweak a prompt mid-session, which for me is half the point of having the prompts at all.

So my server just hands over the prompt text. No middleman execution. Your LLM reads extract_wisdom or analyze_claims directly and stays in one continuous conversation. You can combine patterns or grab just the pieces you need, and your own custom patterns live in ~/.config/custom_patterns/, where they work with any LLM or AI tool that speaks MCP. (You don't even really need Fabric, just any large collection of prompts.)

Why Not Just Fabric?

I could use Fabric as it was initially intended, piping results around like a unix utility. But at this point, I'm just going to pipe to Claude Code, Simon Willison's ahead-of-the-game llm tool, or a context-aware wrapper like plz. Plus, I want to use these prompts with Copilot and Claude Desktop too! And I want to rewrite them on the fly rather than living with a frozen set. Why not think bigger, then? Every knowledge worker is going to need a personal prompt library soon, and writing your own prompts is the fastest way I know to get better at steering these models. You stop retyping the same boilerplate for recurring tasks, and the prompts get sharper every time you touch them. Team knowledge lands somewhere reusable instead of evaporating into a chat scrollback.

Fabric itself is a mixed bag. The curated collection is excellent and the community keeps it current, with CLI tooling that makes direct use pleasant. But the project has strong opinions about which LLM runs your prompt and how. Customizing much of anything means forking. And good luck finding the right pattern when discovery means scrolling through hundreds of files. The documentation is thin on customization. Updates sometimes break custom patterns. There's limited metadata and organization, no built-in versioning or collaboration tools, and plenty of outdated "prompt engineering" stock phrases still baked in. Some super wack prompts that are more like jokes than useful patterns.

The Pattern Audit

So I sat down and read the whole collection, all ~215 patterns, and scored every single one of them. About 32 landed in the A tier, call it 15%. Another 86 or so (40%) are solid B work that a sharpening pass would fix. The C tier holds roughly 65 basic-or-narrow patterns, 30% of the set. And the bottom 32, the last 15%, land in D/F territory: badly built or, in a few cases, flat-out inappropriate.

The security patterns are the best of the bunch. analyze_malware and create_stride_threat_model read like someone who does that work for a living wrote them, and extract_wisdom stays popular for good reason. At the other end you find extract_videoid, which spends an LLM call on what should be a regex, plus constraints like "write exactly 16 words per bullet." Then the basement. find_female_life_partner reduces relationships to algorithms, and no, I'm not making that name up. dialog_with_socrates: seriously, Bill & Ted did it better.

The same problems repeat across the whole collection, and they reveal deep misunderstandings of how modern LLMs work. The IQ claims and virtual hours are cargo cult prompt engineering, mimicking surface features without understanding the mechanism. In the early GPT-3 days, 2020 through 2021, people experimented wildly, and some believed that claiming a high IQ might trigger "smarter" responses by association with high-intelligence training data. There was never evidence this worked. Chain-of-thought prompting, a legitimate technique, got conflated with the pseudo-science, and the pseudo-science spread anyway through copy-paste. "Think step by step" genuinely improves reasoning, and "take a deep breath" is surprisingly effective, probably a quirk of the training data. Role assignment ("You are an expert in X") gives the model useful context. A 1,419 IQ (one pattern claims 24,221) never did anything, and neither did 312 virtual hours of thinking or an imaginary 100m x 100m whiteboard.

Over-rigid constraints come from the same instinct. "Write exactly 16 words per bullet" produces filler like "The deployment process was significantly improved through containerization technology implementation efforts overall." Sixteen words. A person would write "Containerization simplified deployment" and stop. The exact counts grow out of a fear of rambling and some copied academic formatting rules, and the fix is loose guidance, something like "keep bullets concise, typically 10-20 words," with the effort spent on parallel structure instead.

The biggest hole is examples. Around 90% of the patterns don't include a single one, even though one good example settles the edge cases and tone questions that whole paragraphs of instruction leave open. A decent skeleton stays short: the task, a good example, a bad example, and a note on the tricky edge case.

There's also anxiety-driven repetition. One pattern shouts "DO NOT COMPLAIN" three times, which just muddies the prompt. The model was never going to comply harder on round three.

Then there are the ethical misfires. Beyond the relationship-matching disasters, some patterns claim psychological diagnosis capabilities and others invite speculation about strangers' private lives. I'd delete every one of them.

The Pattern MCP Server exists because I got tired of this middleman nonsense, and because I wanted to kick the tires of one of the big prompt libraries out there. You get direct access to the raw prompts, Fabric's collection plus your own custom patterns, inside whatever LLM you're already talking to, and you can hack on them mid-conversation. Reading all 215 patterns convinced me that anyone who works with these models daily needs a prompt library they own and can rewrite. Start with Fabric's collection if that helps. Then rework every pattern you keep until it fits the way you work.