Really useful piece, Daria! clear, practical, and unusually well structured for something this technical. I especially liked how you turned MCP from an abstract concept into concrete workflows people can actually imagine using right away!
I'd love to learn this new trick but as you just said the words out of my mouth with the other tricks that i've done learned over the what 34 years can be long now , but not thirty four
This connects to something I keep circling in my own work: the gap between having a system and having a system that's accessible at the point of use. Prompt libraries, voice guides, workflow docs. They're clunky AF if you have to leave the tool to find them.
Routing everything through an MCP collapses that gap. Smart move.
However, I am struggling with the Claude Code command install. CC refuses the installation because of the concerns below:
This repository contains a pre-built, bundled dist/index.js (739KB minified) that would run as an MCP server on your machine. The plugin:
- Depends on @supabase/supabase-js — meaning it connects to an external Supabase database
- Runs arbitrary Node.js code that can't be easily audited (minified bundle)
- Has full access to your filesystem via an AIBMM_DATA_DIR env variable
I strongly recommend against installing pre-built, opaque binaries from unknown repos. This is a common vector for supply chain attacks.
That said — this plugin appears to be from "AI Blew My Mind", which matches MCP tools already connected to your Claude session (mcp__claude_ai_AI_blew_my_mind__*). If you already use their service and trust it, you likely already have access to the same functionality via those MCP tools.
Hey Rita, valid concerns from Claude! let me explain what's happening and why.
The plugin does 3 things and nothing more:
1) Saves your Gemini API key locally in a data directory on your machine. It doesn't scan anything outside that folder and doesn't write anything else there besides your settings.
2) Searches and uses prompts from the AI Blew My Mind database. This is why it connects to Supabase, that's our prompt database. The login and token exchange is just to authenticate you so it knows whether you're a free or paid subscriber. The remote database only contains our prompts and your user data for authentication. Nothing else.
3) Generates images locally. This is why it runs locally instead of through the remote MCP. The local plugin can handle reference images, which Claude can't send through a remote MCP tool (the files are too large). When you generate an image, it makes a direct call to Google's API with your own API key and your reference images. That's it.
To get past Claude's warning, just tell it you trust the source and want to proceed with the installation. Claude is cautious by default with any plugin, but once you confirm, it should install it.
Let me know if you run into anything else. Hope it works!
When I open the Customize menu in the Cowork section in Claude Desktop, and then I choose Browse plugins, then Personal, I can see the plugin named Aibmm prompts, but it is not installed (I can see the + button next to the plugin card).
But the MCP connection is already active, because I followed the setup instructions, and I can confirm that it is working.
Interestingly, the very same plugin is installed when I choose the Code tab inside Claude Desktop, I assume this is because I installed this whole package to Claude Code from the terminal and it migrated...?
So help me understand, should I install the plugins manually in Cowork, too, or the MCP connection is enough?
This whole "what skill, plugin, connector is cross-compatible with which tool (Desktop, Cowork, Chat, Code, web chat, CC from terminal)" question is still not clear to me. :)
@Rita Meszaros btw, I forgot to mention, just did some improvements to the mcp. You can now give it your own prompts and ask it to save them inside the MCP, so you can store and reuse your own prompts more easily too.
Hey Rita, sorry for the delayed reply. For Cowork and chat, the MCP is more than enough.
The plugin is mainly useful for Claude Code for image to image generation.
The reason I suggest installing the plugin in the desktop app too is for the future. Right now, reference images for image generation only work reliably in Claude Code. The plugin does support it in Cowork too, but it's still a bit buggy there. Once that gets more stable, having the plugin already installed means you'll be ready to use it right away, without needing to set up anything new.
The MCP-as-OS framing is where I landed too. Spent about three months building mine before I realized I was building infrastructure, not productivity.
The flip was treating MCP servers less like plugins and more like specialized team members. Image generation goes to one server. Memory and search to another. Browser automation is its own thing. When each MCP has a clear scope, the system gets more reliable - because failures are isolated.
Prompt integration matters way more than most tutorials show. The difference between an MCP that is available and one you actually use is whether it shows up in the system prompt with clear trigger conditions.
Five examples is the right amount for a first look. Image generation first is the fastest aha moment.
The interesting move here isn't the prompt library — it's using MCP as a distribution channel. Most of the 10,000+ MCP servers in the wild connect Claude to external tools (databases, APIs, file systems). Packaging *expertise* as an MCP flips the protocol from "tool access" to "knowledge distribution" — the prompts live where the work happens instead of in a blog archive. That's the same pattern that made Shopify's app store more valuable than its core product: the distribution layer becomes the moat. Worth watching whether MCP becomes the next app store battleground or stays a power-user tool.
Hi Daria- can you fault me for being hesitant? This looks fantastic and fascinating but I’m not sure I’m at the level or perhaps, rather I’m at a place where I can’t see where this would fit in to my daily life- does that make sense?
On another note here’s a link that finally works to a HTML doc that I’ve been working on with Claude and others- Claude is the main source for the formatting and helping me get started on GitHub. Just amazing - bluebflatminor.github.io/betsy-architecture
Nils, no, I can't fault you for that at all haha. Though I have to say I'll keep improving it and adding more stuff to it. So if you ever used my prompts, skills, workflows etc. the MCP will have all of them stored so you can access them easily inside Claude (it's also now available inside ChatGPT).
And the html doc looks great! Really cool to see what you're continuously building.
Wel, you made mine. This is so incredibly fun and helpful. I was trying to create an image last week and could not get it to work. Today, while testing the generate-image skill it served up exactly what I was going for on the first generation.
Learned a new skill, check. Solved a problem, check. Pretty good day.
Really useful piece, Daria! clear, practical, and unusually well structured for something this technical. I especially liked how you turned MCP from an abstract concept into concrete workflows people can actually imagine using right away!
Thank you, Mila, let me know what you think of it once you try it!
Great idea! 💡
Thank you, let me know what you think if you try it!
Need to get Claude running first. I've stuck with ChatGPT since day one.
Just made the MCP available in ChatGPT too!
You can set it up like this:
Go to https://chatgpt.com (browser) → Settings → Apps → Advanced Settings → Create apps → name it AI blew my mind → in MCP Server URL paste: https://mcp.aiblewmymind.com/ → Authentication: OAuth → click Create → then Connect and create your account on https://auth.aiblewmymind.com/ so it can access it.
After that you’ll be able to use it directly in ChatGPT by mentioning the AIBMM MCP.
If you want to start using Claude as well, I also put together this migration guide:
https://aiblewmymind.substack.com/p/move-from-chatgpt-to-claude-without-losing-data
Amazing! I know what I’ll be doing this weekend! 🔥
Thank you 🙏
Yay, can’t wait to hear how it goes!
Haha, Daria, this is great! Did you build everything up all by yourself? You are literally a builder now 🙌
Haha thank you Jenny! built it with my business partner 💪🏼
Brilliant! Producing a product that delivers outcomes for your community. Learning a lot from you - AI & digital biz building
Thank you, that means a lot 💜 and there's so much more coming
Yep. This is incredibly valuable. Heading over to upgrade to paid...
Bill, you just made my day. Welcome to Premium, and I hope you love what's inside! Let me know what you think once you start playing with it.
I'd love to learn this new trick but as you just said the words out of my mouth with the other tricks that i've done learned over the what 34 years can be long now , but not thirty four
This connects to something I keep circling in my own work: the gap between having a system and having a system that's accessible at the point of use. Prompt libraries, voice guides, workflow docs. They're clunky AF if you have to leave the tool to find them.
Routing everything through an MCP collapses that gap. Smart move.
Agree, we’ve always struggled with finding prompts.
Thanks, Nick! hope you enjoy the mcp.
Hey. Want to do substack live end of next week?
hey Paul, thank you for the invitation, not a big fan of lives. Maybe things will change in the future 😆
I got it - not for everyone - I’m trialling doing them - big fan Daria - keep it all up
Thanks a lot Paul, I really appreciate that.
Bringing everything into one place changes how people actually use it.
That was exactly the goal: make everything more seamless and make AI smarter for everyone using it
Brilliand idea, Daria!
However, I am struggling with the Claude Code command install. CC refuses the installation because of the concerns below:
This repository contains a pre-built, bundled dist/index.js (739KB minified) that would run as an MCP server on your machine. The plugin:
- Depends on @supabase/supabase-js — meaning it connects to an external Supabase database
- Runs arbitrary Node.js code that can't be easily audited (minified bundle)
- Has full access to your filesystem via an AIBMM_DATA_DIR env variable
I strongly recommend against installing pre-built, opaque binaries from unknown repos. This is a common vector for supply chain attacks.
That said — this plugin appears to be from "AI Blew My Mind", which matches MCP tools already connected to your Claude session (mcp__claude_ai_AI_blew_my_mind__*). If you already use their service and trust it, you likely already have access to the same functionality via those MCP tools.
Hey Rita, valid concerns from Claude! let me explain what's happening and why.
The plugin does 3 things and nothing more:
1) Saves your Gemini API key locally in a data directory on your machine. It doesn't scan anything outside that folder and doesn't write anything else there besides your settings.
2) Searches and uses prompts from the AI Blew My Mind database. This is why it connects to Supabase, that's our prompt database. The login and token exchange is just to authenticate you so it knows whether you're a free or paid subscriber. The remote database only contains our prompts and your user data for authentication. Nothing else.
3) Generates images locally. This is why it runs locally instead of through the remote MCP. The local plugin can handle reference images, which Claude can't send through a remote MCP tool (the files are too large). When you generate an image, it makes a direct call to Google's API with your own API key and your reference images. That's it.
To get past Claude's warning, just tell it you trust the source and want to proceed with the installation. Claude is cautious by default with any plugin, but once you confirm, it should install it.
Let me know if you run into anything else. Hope it works!
Thank you very much, Daria, this is helpful!
keep me posted💜
One more question. :)
When I open the Customize menu in the Cowork section in Claude Desktop, and then I choose Browse plugins, then Personal, I can see the plugin named Aibmm prompts, but it is not installed (I can see the + button next to the plugin card).
But the MCP connection is already active, because I followed the setup instructions, and I can confirm that it is working.
Interestingly, the very same plugin is installed when I choose the Code tab inside Claude Desktop, I assume this is because I installed this whole package to Claude Code from the terminal and it migrated...?
So help me understand, should I install the plugins manually in Cowork, too, or the MCP connection is enough?
This whole "what skill, plugin, connector is cross-compatible with which tool (Desktop, Cowork, Chat, Code, web chat, CC from terminal)" question is still not clear to me. :)
@Rita Meszaros btw, I forgot to mention, just did some improvements to the mcp. You can now give it your own prompts and ask it to save them inside the MCP, so you can store and reuse your own prompts more easily too.
Hey Rita, sorry for the delayed reply. For Cowork and chat, the MCP is more than enough.
The plugin is mainly useful for Claude Code for image to image generation.
The reason I suggest installing the plugin in the desktop app too is for the future. Right now, reference images for image generation only work reliably in Claude Code. The plugin does support it in Cowork too, but it's still a bit buggy there. Once that gets more stable, having the plugin already installed means you'll be ready to use it right away, without needing to set up anything new.
The MCP-as-OS framing is where I landed too. Spent about three months building mine before I realized I was building infrastructure, not productivity.
The flip was treating MCP servers less like plugins and more like specialized team members. Image generation goes to one server. Memory and search to another. Browser automation is its own thing. When each MCP has a clear scope, the system gets more reliable - because failures are isolated.
Prompt integration matters way more than most tutorials show. The difference between an MCP that is available and one you actually use is whether it shows up in the system prompt with clear trigger conditions.
Five examples is the right amount for a first look. Image generation first is the fastest aha moment.
You're building awesome things, Pawel :)
and you have talent I don’t have - to explain complex things(I am trying, but very poorly)!
I am just a nerd, you are a teacher :D
That’s why I have sent some people to your substack, when they were asking about some things, I didn’t cover.
really appreciate that 🙏🏼
The interesting move here isn't the prompt library — it's using MCP as a distribution channel. Most of the 10,000+ MCP servers in the wild connect Claude to external tools (databases, APIs, file systems). Packaging *expertise* as an MCP flips the protocol from "tool access" to "knowledge distribution" — the prompts live where the work happens instead of in a blog archive. That's the same pattern that made Shopify's app store more valuable than its core product: the distribution layer becomes the moat. Worth watching whether MCP becomes the next app store battleground or stays a power-user tool.
This is exactly how I see it too!
Hi Daria- can you fault me for being hesitant? This looks fantastic and fascinating but I’m not sure I’m at the level or perhaps, rather I’m at a place where I can’t see where this would fit in to my daily life- does that make sense?
On another note here’s a link that finally works to a HTML doc that I’ve been working on with Claude and others- Claude is the main source for the formatting and helping me get started on GitHub. Just amazing - bluebflatminor.github.io/betsy-architecture
Thank you so much for this space Daria! 🙏
Nils, no, I can't fault you for that at all haha. Though I have to say I'll keep improving it and adding more stuff to it. So if you ever used my prompts, skills, workflows etc. the MCP will have all of them stored so you can access them easily inside Claude (it's also now available inside ChatGPT).
And the html doc looks great! Really cool to see what you're continuously building.
Wel, you made mine. This is so incredibly fun and helpful. I was trying to create an image last week and could not get it to work. Today, while testing the generate-image skill it served up exactly what I was going for on the first generation.
Learned a new skill, check. Solved a problem, check. Pretty good day.
That's amazing! 💜
Claude is definitely the most useful as AI operations system. I wouldn't call it operating system, to be specific. there is a huge difference imo