Tell It Once: Setting Up Claude With Skills and MCP Servers

2 minutes ago

12 min read

Many of us use AI the same way every single day. Open a tab. Paste in a question. Copy the answer out. Fix what it got wrong. Then tomorrow, we start over, because it remembers nothing.

That was the opening of my session at JaxDUG on Thursday night, "Intro to Claude AI: Your Efficiency Assistant." Our scheduled speaker had a last-minute conflict, so I filled the slot with the tool I use every day.

When I polled the room, we had the full spread. Some people use AI every day. A couple tried Copilot early, gave up, and came back later. One attendee's company blocks the Claude website entirely. Wherever you land on that list, this post is for you.

As promised, here is the link to the slides: bit.ly/IntroToClaudeAI

Thank you to Keiser University for hosting us, and to JaxData for sponsoring the night.

What It Doesn't Know

The problem usually isn't the model. It's that the model doesn't know anything about you:

Your voice: how you actually write
Your standards: naming conventions, structure, the patterns your company expects
Your tools: your database, your meeting notes, your files
Your history: what you decided last week, and why

So you re-explain it all every time. The fix comes in three layers: point it at your work, write your standards down once (skills), and plug it into your actual tools (MCP servers).

One thing to note: none of this trains the model. Your skills and files are instructions that ride along with your prompts, the same ones you'd otherwise type every single time.

Setup

Claude runs in three places, and it's the same model behind all of them:

Browser (claude.ai): quick questions and research. In my opinion, too many people use AI as a glorified Google. If a search finds the answer, use the search.
Terminal (Claude Code): files, code, scripts, and long-running projects.
Desktop app: the one I use most, because it pulls everything into one place.

Installing Claude Code is one line, then point it at a folder:

npm install -g @anthropic-ai/claude-code
cd C:\ClaudeCo-Work
claude

What does it cost? I started on the $20/month plan. Then I built a tool to automate my SQL Server health check, which I usually do manually, and I kept hitting my usage limit after a question/adjustment or two a day because the project got so large. I ended up moving up to Max at $100 a month. Here's how I look at it: that's less than one billable hour. Start on the lower plan and move up when you actually hit the wall. It's worth every penny if you use it and configure it correctly.

One Folder for Everything

The first thing I did was create one working folder: C:\ClaudeCo-Work. To give a little bit of background, Claude was saving finished files to a temp path so long that Windows wouldn't open them (thank you, Windows 255 character path limit).

Now everything lives in one short path, I always know where my work is, and I can restrict Claude's permissions to that one folder.

CLAUDE.md

Inside that folder is CLAUDE.md. It's the first thing Claude reads, like a startup batch file for your assistant. Here's mine:

# Jeff Taylor's working folder

Read these before doing anything on Jeff's behalf. They already exist.
Never ask Jeff to re-explain his style or preferences, and never say
you can't reach them.

- writing-style.md: voice and habits for anything written as Jeff. No em-dashes. Ever.
- presentation-style.md: slide decks and presentations.
- claude-preferences.md: file saving, platform limits.

## Standing rules

- Save all final deliverables directly to C:\ClaudeCo-Work.
- If a target file is locked (open in Excel or Word), save a _v2 copy rather than failing.
- Jeff is premium on LinkedIn (3,000 chars) and X (25,000). Do not cap X posts at 280.

Notice that every rule in there exists because something went wrong. It kept asking where my style file was, even though it was sitting right there. It failed every time I had a spreadsheet open. It kept writing me 280 character posts for X. I got tired of saying the same things, so I wrote them down.

Now it's a text file. That's the whole trick.

Skills

A skill is a folder with a SKILL.md file that explains how you do a thing. That's it. You explain it once, and it applies in every session after that.

Since they're just text files, you can back them up and share them. One gap I've run into: my setup doesn't sync between my work laptop and my personal laptop. My plan is a GitHub repo that both machines pull from.

My Voice, Written Down

I didn't want to tell Claude how I write every time. So I pointed it at my blog posts on my website and said: go read these, figure out my tone, punctuation, and structure, and write me a style guide.

It did. The result is writing-style.md, built from my posts. Out of curiosity, I also asked how my writing has changed since I started blogging. The answer: the voice has been consistent the whole time. "Here's the problem; here's how to fix it, step by step."

A rule I care about... No em-dashes. Ever. AI loves them. I don't use them. If my name is on it, it needs to sound like me, and I still read every word before its published.

I mentioned in the session that my style file is 16KB, and I wondered if that was too big. I went and checked. Anthropic's skill authoring best practices recommend keeping the body of a SKILL.md under 500 lines and a max of 5000 words, because once it loads, every token competes with the rest of your conversation. My style file is about 220 lines, so I'm fine for now. Still, keep them lean.

presentation-style.md does the same thing for my slides: a cold open that gets right to the point, fragment reveals (a slides.com term similar to animations in PowerPoint) instead of full slide transitions (one bullet at a time, so you listen to me instead of reading ahead), styling at the theme level so there's one place to change it, and an evaluation close. When I speak at a new conference, I hand Claude the event's logo and colors and say, "theme it like this."

My Own Skills

I wrote four skills that I use:

presentation-theming: turn a brand into a slide theme
slides-com-deck-authoring: how my decks get built
web-development: how I want web code written
mobile-development: my .NET MAUI conventions

For the development skills, it stores my architectural preferences. Naming conventions. Security. Should the API be rate-limited? Caching with Redis? Dapper or Entity Framework? Those are the questions I work through before I design any application, so I wrote them down. When I ask Claude to build something, it already knows how I'd architect it and can focus on the functionality. A skill can also say, "When you reach this decision, ask me which option I want."

The mobile skill is getting a workout right now. My rule for my kids: if you want a phone or a computer, you have to build an app first, so you understand what it takes (as I did back in the day). My oldest daughter and I are building ours in .NET MAUI, and the skill carries the things you don't think about until later, like retrying API calls with no connection and syncing local data when the connection comes back.

Skills I Didn't Write

I didn't write most of the skills I use. Some ship with Claude: Word, Excel, PowerPoint, PDF. Some arrive as plugins, a whole bundle at a time: Slack, Canva, Zapier, Wix, Auth0, Postman. There are role packs too, for engineering, data, and operations. I've used those to review my own work from someone else's frame of reference: "If I'm an executive, reading this summary, is this what I'm looking for?"

Skills vs Files

A skill announces itself and loads when the work matches. A file just sits there until something goes and reads it. Both work. Only one is automatic.

If you keep repeating yourself, promote the file to a skill.

The Easy Way to Make One

You can write the markdown yourself, but that's not what I do. I do the work first. Then, when we're done, I tell Claude:

We just did all this work. Take everything you've learned from what I've taught you and create a skill for it.

The next time I work on a deck, I ask: "Is there anything new we did here that isn't in the skill yet?" If there is, it updates the skill. Same with my writing style. If it writes something I would never say, I tell it to remove it and fix the rule.

Every correction makes the skill better, so you review less, iterate less, and send fewer tokens for the same result. That's also why I haven't come close to my usage limit since.

Memory

Claude now keeps notes on how I work across every session. I corrected it once on how I fragment a slide, and I haven't had to say it a second time.

Tell it once. That's the whole job.

MCP Servers

MCP (Model Context Protocol) is a standard way to plug a tool into the assistant. The old way: copy data out, paste it in, copy the answer back. The new way: it reads the source itself.

Here's what I have connected: SQL Server, Granola (my meeting notes), slides.com, Wix (my blog), Bitly, Slack, and Postman.

Demo 1: A Blog Post in My Voice

For the first demo during my presentation the other day, I asked the room for a blog topic.

One attendee suggested Shopify's announcement that it's moving its mobile apps off React Native and back to native.

Granola was recording the session, so I asked it for a paragraph on what he'd just suggested, pasted that into Claude, and asked it to write a blog post. You could watch it work: it loaded my writing style, read my preferences, then went and read Shopify's engineering posts. By the time we finished the next demo, it had written about 1,700 words, saved the markdown to my folder, and pushed a draft to my Wix blog.

Was it perfect? No. The formatting in Wix needed some spacing, and it needed an image. And I'm reading every word before it goes live. But the heavy lifting was done.

However, there is a catch: a model doesn't know what happened after it was trained. I ran into this with SQL Server 2025. Claude kept writing a pile of code to solve something a new 2025 function already handles, because it had never seen the documentation. I told it to read Microsoft's SQL Server 2025 docs first, and then it got it right. Point it at current sources, and check what it brings back.

Demo 2: Ask SQL Server a Question in English

This is the one a room full of developers and DBAs leans in for. The SQL Server MCP server can list databases, read a table schema, list stored procedures, run a query, and read a stored procedure definition.

Before I connected anything, I created a dedicated login for Claude with access to one database, read-only. Here is a sample user creation script similar to what I wrote:

USE [master];
GO
CREATE LOGIN [claude_mcp] WITH PASSWORD = N'CHANGE_THIS_PASSWORD';
GO
USE [YOUR_DATABASE];
GO
CREATE USER [claude_mcp] FOR LOGIN [claude_mcp];
ALTER ROLE [db_datareader] ADD MEMBER [claude_mcp];
GRANT VIEW DEFINITION TO [claude_mcp];
GO

Then I asked: "Give me a list of all databases in my local SQL Server, then list all the tables."

It found 12 databases (4 system, 8 user) and told me it could only list tables in one of them, because its login has no access to the other 7. In the one it could see, it listed 107 tables across the dimension and fact schemas, plus 37 staging tables.

Then it helpfully suggested I could run it as sysadmin to see the rest.... No thanks! We are not doing that.

To configure the MCP for SQL Server, you need to set up a small block of JSON code in your Claude config: a name, a command, arguments, and a connection string. Add a separate entry for each server you want to reach.

{
  "mcpServers": {
    "mssql": {
      "command": "npx",
      "args": ["-y", "YOUR_MSSQL_MCP_SERVER_PACKAGE"],
      "env": { "MSSQL_CONNECTION_STRING": "CHANGE THIS" }
    }
  }
}

The package name and settings depend on which SQL Server MCP server you choose, so read its docs before you install it.

Here's the part I don't want anyone skipping:

It runs as whoever you connected it as
Use a read-only account for demos and exploring
It will happily run whatever you ask it to run
Never point it at production on a whim. Generate the scripts, run them in Dev and QA, then promote them like any other change.

Granola: My Meeting Notes

Granola captures the audio on my PC or phone, transcribes it, and summarizes it. I can ask "what did I commit to on that call?" or have it write the follow-up email.

With the MCP connection, it opens more opportunities. When working with clients over time, with many meetings, status updates, etc., I can ask Claude to build a plan, task list, or documents. It reads every meeting, pulls in documents and any additional data from SQL Server if needed, and produces a detailed plan: what's done, what's still open, and where the risk is.

Fun fact: this session was recorded in Granola too. That transcript is where this post started.

Demo 3: The Countdown Timer

A real problem I had recently was that I needed a count down timer that would count down from various times and then also rotate slides or a full page image for our breaks for a Dashboard in a Day for Power BI training I was giving.

PowerPoint can't count down on its own. Claude's first try was a VBA macro, which is blocked on a lot of machines. Its second try was a deck with 452 slides, one for every tick of the clock.

452 slides. That was a bit much, I am not editing that.

So I went a different way: rotate 4 images in a local HTML page every 8 seconds with a live countdown, and let me save a copy as a 45 minute or 10 minute version. That worked. Then I asked Claude to write a prompt capturing everything we'd built. The prompt it wrote was far more detailed than anything I'd typed: preset times, drag and drop images sorted so slide#2 comes before slide#10, keyboard shortcuts, and a 'Save a copy' button that bakes the images and time into a file that starts the moment you open it. One HTML file, no internet, no external libraries.

My Presentation Deck

The deck I use for my session the other day was built the same way. Claude built the slides from my presentation rules, skinned them in the JaxDUG palette pulled from the logo, pushed them to slides.com over MCP, then rendered every slide and checked its own work.

To be clear, it wasn't one clean prompt. I corrected a few things (the reveals/transtions on the demo slides, for one), and every correction went back into the skill.

Task	Before	After
Blog post draft	2 hours	About 20 minutes
Re-skin a deck for a venue	An evening	One CSS block

Keep in mind the "after" column still includes me checking the work. For one conference, I had Claude recreate the curves from the event's logo as the slide background, all in CSS. I know CSS pretty well, and that still would have taken my whole evening.

Q&A From My Presentation

When I presented at JaxDUG the the other day, we had a few questions...

GitHub Copilot or Claude? It depends on the model more than the tool, but I prefer Claude's interface. I built a C# weather API entirely in Claude. I didn't review diffs in a side panel. I ran it, tested it, and stepped through the debugger. That's where I caught a couple ologic bugsic. I pointed them out, it agreed, and it fixed them.

How do I keep token usage down? Skills and MCP servers first. Second, a long chat sends its whole history with every message. When a chat gets huge, ask it to "write me a prompt for everything we've done so far" and start fresh with that. You lose some context, so I don't do it every time.

Is my data training the model? Check your settings. There's a privacy option that controls whether your chats and coding sessions can be used to train future models. My company had us turn it off to keep client information out, and I did the same.

A tip from the audience: tell it to record every mistake it makes, so it never makes the same one twice. I haven't tried that yet. I will.

Conclusion

If you take nothing else from the session, take the checklist:

Use existing skills. You don't have to build anything to get value out of day one.
Write your own. Start with the thing you explain multiple times. A text file with your standards in it changes the output more than any clever prompt.
Let it generate the skill for you. Do the work first, then say "take everything you just learned and write me a skill for it." That is how all four of mine got written.
Install a plugin. Slack, Canva, Zapier, Wix, Auth0, Postman, and the role packs for engineering, data, and operations. Somebody already did the work, so use it.
Plug in one MCP server. Pick the tool you copy and paste out of the most. For me, it was SQL Server. Read-only, one database, and check every query it writes.
Turn on memory and correct it. Tell it once, and the correction sticks across every session after that. Fold the good ones back into your skills.

You don't have to do all of this to see value. Do one, and you'll see the difference the next time you ask for something.

AI is an assistant, not a replacement for knowing your craft. I still read every line of T-SQL and every blog or document it writes, and so should you. It is only as good as what you have taught it, and everything I showed at JaxDUG came out of getting tired of repeating myself. Configure it, correct it, and keep the corrections.

Want to dig in? Start with the Claude docs at docs.claude.com and modelcontextprotocol.io. The session slides are at bit.ly/IntroToClaudeAI.

If you try any of this, I'd love to hear what you build. Leave a comment below or reach out!