Kyle
banner
kylestratis.com
Kyle
@kylestratis.com
📊 Data, AI/ML, and Cloud consultant @ Stratis Data Labs
🚉 VPEng @ Sequel Institute
🤓 Startup advisor
✒️ Writing about MCP
🐍 Python
☁️ AWS
🧙‍♂️ Elixir
🐷Ham radio
🇺🇸↔🇬🇷 [Boston // Athens]
Mostly tech things here
I felt vindicated when I read @joereis.bsky.social's newsletter this morning highlighting this (still early) trend and speculating about where it might be going.
November 9, 2025 at 8:21 PM
I've gotten into arguments about this elsewhere, but this is where I see lots of software going. At the business level build vs buy is about to be turned on its head, personally almost anyone can build the exact (or close enough) software they need.
November 9, 2025 at 8:21 PM
I built a Spotify MCP server for an idea I’ve been noodling on for using natural language to make one-off playlists based on several different similarity measures or even “make a playlist of tracks from these 3 artists” so I might hack on that or the app layer a bit if I can find the time
November 8, 2025 at 12:11 AM
Eternally working on my book, it feels like. Somehow I have to get 3 more 10kish word chapters done by the end of the year 🙃
November 8, 2025 at 12:09 AM
Off the top of my head Cursor allows you to do this, I think Claude Desktop does but I don’t use it too much with MCP, and maybe GitHub copilot? My understanding is that most of the agentic IDEs and extensions like Cline do as well.
November 6, 2025 at 3:47 PM
Most of the major clients have that built in. If you’re building something with agents you should build that as well according to your user interaction model. Most servers have way too many tools on their own to begin with. GitHub’s has like 90 last I checked.
November 6, 2025 at 2:04 PM
I was similarly confused reading this thread.
November 4, 2025 at 6:52 PM
Now that I’m recovered from both the talk and Halloween, it went really well! It helped that I couldn’t see the audience 😅
November 1, 2025 at 12:40 PM
🙏🙏🙏
October 30, 2025 at 7:10 PM
The distribution use cases are different as well, and my god skills chomp up lots of tokens, while servers are generally more efficient, mostly because in practice they focus on tools.
October 30, 2025 at 5:11 PM
I don’t think so. Even where there is overlap (providing prompts), in Skills prompts are model-controlled, in MCP they are not. Skills are limited to Anthropic, MCP is universal. Even the way scripts are provided looks different to how tools are, and they’re largely controlled by the prompts.
October 30, 2025 at 5:11 PM
(Spoiler alert! 😉)
October 28, 2025 at 10:15 PM
It’s so good, h/t to @randyau.com for getting it going!
October 28, 2025 at 3:33 PM
So I started a project that's been on the backburner, fired up Claude Code, and went against my own instincts by vibecoding an MCP server for a project I've been wanting to build for a while. You can read about how it went below:
kylestratis.com/posts/payin...
October 28, 2025 at 1:36 PM
The model can decide whether it needs to use the Skill, much like a tool, and activates it as needed.

What's kind of brilliant is that the model only uses the main skill - anything bundled with the skill also only gets loaded a need-to-use basis, attempting to save you on token costs.
October 28, 2025 at 1:36 PM
Skills are the latest feature released by Anthropic, allowing you to essentially give Claude Desktop, Claude Code, Claude Apps, or applications using the Claude API onboarding docs, providing them direction and guidance for specific tasks.
October 28, 2025 at 1:36 PM