PowerUp Coaching — Living Knowledge Base

Insight: Build a Living Prompt Library Behind an MCP Server

7 min read

Original Insight

“I have all my prompts in here, and then I can put this behind an MCP. Anytime I want, I could just go in, edit the prompt, bring it up to date, and it’s instantly available to all of my applications that use that prompt. In 7 minutes, I have this thing public, online, and ready to go. I didn’t have to set up a tech stack. Literally one paragraph of input.” — Lou

Expanded Synthesis

The most overlooked bottleneck in an AI-powered knowledge business is not model selection, context engineering, or even workflow automation. It is prompt versioning. Every time you improve a prompt — whether it is your newsletter rewrite assistant, your coaching intake analyzer, your content ideation engine, or your client onboarding sequence — you face a fragmentation problem: that improved prompt exists in one chat window, or one saved text file, or one browser bookmark, but all the agents and automations built on the previous version are still running the old one.

The insight Lou demonstrated in this session is elegant precisely because it addresses that fragmentation at the infrastructure level. A prompt library with a CRUD interface and an MCP (Model Context Protocol) backend is not a novelty project — it is the foundation of a scalable AI-assisted practice. When your prompts live in a queryable, versioned database that is accessible via an API that any agent can call, you have separated the prompt as an asset from the specific context in which it is used. You can improve the prompt once and every workflow that depends on it is immediately upgraded.

The Cloudflare Workers deployment Lou used to build this in seven minutes is worth examining for what it represents beyond the specific technology. It is an example of the threshold being crossed where the time cost of building infrastructure has collapsed to near-zero. A year ago, setting up a prompt management application with a UI, a database, and an API endpoint would have taken days of skilled development time. Today it takes a paragraph and seven minutes. The implication is not just that it is faster — it is that the excuse of “I’m not technical enough to build that” has largely expired.

What separates this from simply saving prompts in Notion or a Google Doc is the MCP connection. When a prompt is accessible via MCP, your AI agents can pull prompts by ID, by tag, by category, or by query — without you having to manually copy and paste anything. An automation in N8N or Make.com can be designed to always fetch the latest version of a prompt before executing. A Claude Code session can reference your prompt library as a tool call. The prompt becomes a living, managed asset rather than a static artifact.

The versioning dimension is critical and often underappreciated. Lou noted that the OpenAI Playground now provides prompt versioning with comparison views, performance evaluation, and the ability to call a specific prompt version by its ID from the API. This means that A/B testing your prompts — which most sophisticated operators do informally (“this version feels better”) — can become rigorous. You can run the same input through prompt_v1 and prompt_v2, evaluate outputs against a rubric, and make data-informed decisions about which direction to refine.

For PowerUp clients who are building their coaching practice on AI, this insight resolves a tension that often emerges once you have built several AI-assisted workflows: the fragmentation between what you’re still learning and what you’re already relying on. When you are improving your coaching intake prompt, you don’t want to break the automation that processes new client forms. When you are refining your content generation system, you want the new version to propagate to all your publishing workflows without manual reconnection. The living prompt library solves this problem structurally.

The deeper coaching implication is about what happens to your intellectual property when you codify it this way. Your prompt library is, in a meaningful sense, a map of your methodology. The way you have written prompts for client discovery, for assessment interpretation, for framework delivery, for follow-up — that is your coaching system externalized. A well-maintained prompt library is not just an operational tool. It is a proprietary methodology that can be sold, licensed, trained on, or handed to an AI employee. It is the difference between having a practice that runs through you and having a practice that runs because of you.

Practical Application for PowerUp Clients

The Prompt Asset Inventory

Before building a prompt library, you need to know what you have. Complete this inventory:

  1. List every repeated AI task you currently perform — things you do more than once a week with AI. Examples: drafting client emails, creating session summaries, generating content, analyzing client feedback, repurposing posts, refining frameworks.

  2. For each task, identify whether you have a prompt — a consistent instruction set you use every time — or whether you are writing it fresh each time.

  3. Tag each prompt by category: Client Work, Content, Research, Automation, Personal.

  4. Identify your top 5 “anchor prompts” — the ones your practice could not function without. These are the prompts that, if lost, would cost you the most to reconstruct.

  5. Build your first prompt library. Use Cloudflare Workers (build.cloudflare.dev) as Lou demonstrated for a zero-cost, immediate deployment. Or use Notion with a database view if you prefer simplicity over API access. The discipline of centralization matters more than the specific platform at this stage.

The Prompt Update Ritual: Once a month, review your top 5 anchor prompts. For each one, ask:

  • Does this still reflect how I think about this task?
  • Has the model I use for this changed in a way that requires updating?
  • Is there a pattern from my last 30 days of work that should be encoded here?

Update, version, and note the date and reasoning for the change.

Coaching Questions:

  • What is your most valuable prompt — the one that encodes the most of your expertise? Does it exist anywhere outside of a chat window?
  • If you were to hire an AI assistant to run your practice for a week, what prompts would you give them? Have you written those down?
  • What would it be worth to you to have every AI tool you use always running your best, most current thinking rather than whatever version you last remembered to copy and paste?

Additional Resources

Evolution Across Sessions

This insight is the operational implementation of “Codify Your Judgment Into Skills, Not Just Prompts.” Where that earlier insight argued for the principle of encoding your methodology, this one provides the infrastructure: a versioned, queryable, MCP-accessible prompt library is the technical form that encoded judgment takes. The multi-agent context manager Lou also discussed in this session (routing queries to Grok, GPT, and Claude in parallel, integrating responses, and re-circulating context) depends on this kind of prompt infrastructure to be manageable at scale.

Next Actions

  • For me (Lou): Share the Cloudflare Workers prompt library demo link in Telegram. Work through the MCP URL issue that prevented Claude Desktop connection. Build out the N8N multi-agent debate workflow promised for the Oct 9 session.
  • For clients: Complete the Prompt Asset Inventory above. Set a target of having your top 5 anchor prompts in a single centralized location (any platform) within two weeks. Note which ones are currently only in chat history — those are at risk.

Graph View

Backlinks