The first time a meal kit lands on your porch, it feels like one thing. One box, one dinner, one charge on the card. Unpack it and it's really four things riding together. There's a recipe card, which is the method. There's the food, which came from suppliers you never deal with yourself. There's your own kitchen, with the salt, the oil and the pan you like, which the box quietly assumes you already have. And there's the box, which exists so all of that shows up at once.
Once you see it that way, the recipe card is the part worth keeping. You can make that dinner again next month with groceries from anywhere, in anybody's kitchen. The box goes in the recycling.
A custom GPT was one box too. Instructions, reference files, standing facts about your agency and the occasional link to another system, all sealed in together. OpenAI retires custom GPTs on December 11, and Google begins turning Gems into skills in November for personal accounts. Every major vendor is now unpacking the box the same way, into four pieces: skills, projects, connectors and plugins. The labels shift a little from vendor to vendor, but the shape underneath holds, and once you know it, the question of where your agency's work should live gets a lot easier to answer.
A skill is the recipe card
A skill is a written procedure for one job. How to build a quote comparison for a prospect, how to format a renewal summary, what to check before a certificate goes out the door. It's saved as a folder with a file called SKILL.md at its center, plus any templates or examples that belong with it.
The AI doesn't read every skill all the time. It keeps a short description of each one on hand and pulls in the full instructions only when a request calls for them. Anthropic's help center says skills are task-specific and "only load when relevant," and Microsoft describes them in Copilot Studio as "reusable task-specific capabilities."
Skills are also the one piece that travels. The format is an open specification published at agentskills.io, OpenAI and Microsoft both describe skills as reusable instructions for a specific task, and Google's help pages say a SKILL.md made on another platform can be uploaded into Gemini. A skill that's mostly well-written instructions can move with you the next time a vendor retires something.
If you'd write it on an index card and hand it to a new hire, it's a skill.
A project is your kitchen
A project is a workspace with standing instructions and files that are there every time you start a conversation inside it. Who the agency is, which carriers you're appointed with, how client emails should sound, that nothing in writing ever promises coverage, and this quarter's renewal list. Anthropic describes Projects as "static background knowledge that's always loaded," which is exactly the job. ChatGPT offers Projects on every plan.
Google's nearest equivalent is a notebook. A NotebookLM notebook holds your sources and custom instructions, now shows up inside the Gemini app as well, and in NotebookLM itself it answers only from the sources you've loaded. That makes it a good home for carrier underwriting guides and policy wordings, where you want answers pulled from the document rather than improvised around it.
Projects don't travel, and that's fine, because what's in them is yours. It's also where client files belong. A skill is built to be copied and shared, so client details stay in the project with the work. Anything that has to be true in every conversation goes here.
A connector is the supplier
A connector is an authenticated link to a system you already use, like your email, your drive, your calendar or your CRM, so the AI can read from it and, where you allow it, act in it. In Claude, every connector is a connection to an MCP server. MCP, the Model Context Protocol, is an open standard, and its own documentation compares it to a USB-C port for AI applications. OpenAI calls these apps, Microsoft calls them tools, and Anthropic calls them connectors. It's the same job under three names, and the test for it is simple: if it reaches outside the chat window, it's a connector.
The line between a connector and a skill is the most useful one on this page. Anthropic's help center puts it this way: connections give Claude access to tools, while skills teach Claude how to use those tools. A connector to your management system without a skill is a new hire with a login and no training. A skill without the connector is the training with no login.
A plugin is the box
A plugin packages skills and connectors, and in some tools commands and helper agents, into one install. Anthropic's documentation says a plugin for a product "bundles that product's connector with the skills that use it." OpenAI's plugins do the same for ChatGPT, and when you migrate a custom GPT, the button OpenAI gives you says Migrate to plugin.
You don't need a plugin to use a skill or a connector. Both work on their own. A plugin earns its place when several people need the same setup on the same day, and on Claude's Team and Enterprise plans an owner can install one for everyone. Build the pieces first, and box them up when someone else needs the whole setup.
All of it on one page
| Piece | What it holds | When the AI reads it | Travels to other tools | Agency example |
|---|---|---|---|---|
| Skill | A written procedure, with its templates and examples | Only when the task calls for it | Yes, it's an open format | Your quote comparison format |
| Project | Standing instructions and files | Every conversation inside it | No | Agency tone, carrier list, this quarter's renewals |
| Notebook (NotebookLM) | Your source documents | Answers grounded in those sources | No | Carrier underwriting guides |
| Connector | Access to an outside system | When the AI needs to read or act there | The protocol is open; each connection is per system | Email, drive, calendar |
| Plugin | Skills and connectors in one package | Installed once, then like its parts | Within its own platform | Your team's standard setup |
| Custom GPT or Gem | All of the above in one box | Always, as one unit | No | Retiring: GPTs on December 11, 2026; Gems from November 2026 on personal accounts |
One agency job, sorted
Say your producers spend Tuesday mornings turning carrier quotes into a side-by-side comparison for prospects. Here's where each part of that job lives.
The comparison itself, meaning which columns, what order, how to flag a coverage gap and how to word the cover note, is a skill, along with a finished example or two as reference files. The agency's name, the carriers you write with, the tone you want and the rule that nothing in writing promises coverage go in the project, because they're true whatever the job. The carriers' underwriting guides go in a notebook, or in the project's files if you're working in Claude or ChatGPT.
Pulling the quotes out of email and saving the finished comparison to the client folder takes connectors to your email and your drive. Your management system only belongs on that list if its vendor offers a connector, so ask before you assume one exists. And once three producers are running it the same way, the skill and its connectors are worth boxing up as a plugin so the next hire gets the same setup on day one.
That's also roughly how I run my own Digital Crew: procedures as skills, standing context in projects, and connectors doing the reaching. How I Use Notion MCP to Give My AI Crew a Permanent Memory shows the connector side of that in practice.
Where the edges are still rough
The four pieces are converging, and they haven't finished. Google says Gems will turn into skills with sharing and Drive files carried over, and it describes some of that as coming rather than here. When I tested it in September, a Gemini skill I'd made in Spark could be called into any Gemini chat or notebook, but not into a notebook opened in the older NotebookLM interface. That will probably change, which is exactly why it's worth checking on the day you build rather than trusting a post, this one included.
The vocabulary drifts too. Microsoft's Copilot Studio splits the project idea into instructions and knowledge. OpenAI calls connectors apps. None of that changes the sorting. Whatever the vendor calls the pieces, you're still asking the same questions about each part of a job.
Where to start
The next time someone at your agency says "let's build a GPT for that," sort the job before anyone opens a builder. Whatever has to be true in every conversation goes in a project. Anything with steps and a finish line becomes a skill. Whatever reaches into another system needs a connector, and a plugin waits until someone else needs the whole setup. That sorting tells you what to build, and the skill you end up with will outlast whichever tool you're using this year.
If you've got a few GPTs to move, How to Migrate a Custom GPT to a Skill walks through the sorting one GPT at a time, and OpenAI Is Retiring Custom GPTs. Here's the Official Record. has the dates. For the Claude side in more depth, How to Turn Claude into a Digital Crew (with Skills and MCP) picks up from here.
If you'd like a clearer picture of where your agency stands with AI before you rebuild anything, the AI Adoption Reality Check takes about five minutes. I read every one and write back within two days.
Frequently asked questions
Is a Claude Project the same as a ChatGPT Project? Close enough to plan with. Both hold standing instructions and files for the work you do inside them, and neither one moves between vendors, so switching means pasting your instructions into the new one and uploading the files again.
Do I need a connector to use a skill? No. A skill that formats or checks work you hand it needs no connection at all. It only needs a connector when the job means reading from or writing to another system.
What happens to my custom GPT or Gem? OpenAI's migration turns a custom GPT into a skill inside a plugin, where your plan supports skills, and GPTs stop working on December 11, 2026. Google says it will turn Gems into skills automatically, starting in November 2026 for personal accounts and March 2027 for Workspace business accounts.
Which piece should an agency set up first? A project. It's the quickest to build, it stops the re-explaining of who you are in every chat, and it gives your first skill a place to run.
The Skills series
Every post in the series on custom GPTs, Gems and skills, in the order it ran.
- OpenAI Is Retiring Custom GPTs. Here's the Official Record.
- How to Migrate a Custom GPT to a Skill (ChatGPT or Claude)
- On ChatGPT Plus? Your Custom GPT Options After December 11
- Skills vs Projects vs Connectors vs Plugins, Explained for Agency Owners (you're here)
- 8 AI Skills Every Insurance Agency Should Build First
- Your Medicare AI Prompts Are Still Running Last Year's Rules
