So What? Growing Prompts Into AI Skills

Katie Robbert – 00:33

Happy Thursday. Welcome to So What? The Marketing Analytics and Insights Live Show. I’m Katie, joined by Chris and John. Hello, fellows. This week, we are talking about growing prompts into AI skills.

Three-ish plus years ago, when generative AI hit the public market for the general consumer, everyone was struggling to try to figure out how to prompt correctly, effectively, and efficiently. That’s where things like the Trust Insights REPEL framework were born. We’ve had image prompts, we’ve had RACE, and we’ve had PAIR. There are a lot of different frameworks for how to do effective prompts.

However, now you have systems that can build in skills and take all of your prompts and automate them. The transition was having a general prompt structure that you paste in every time. Then you build a custom GPT, a Gem, or a project—depending on the system you’re in—and those have your prompts baked in so you don’t have to copy and paste them over and over again.

The next evolution of that is: how do you take that custom thing and make it a skill that you can apply across anything you’re doing so you don’t just have to work within that custom Gem, GPT, or product? That’s what Chris is going to walk us through today. By “us,” I mean me, John, and whoever is watching.

Christopher Penn – 02:24

Let’s start with a bit of table setting on skills. Skills is a specific term—Skills with a capital S. This is a data format that Anthropic, the makers of Claude, originally came up with for the Claude ecosystem. The best way to think about it is like a GPT or a Gem that you can move anywhere.

One of the biggest criticisms of things like custom GPTs was that you have to use ChatGPT with it. If you have a Gem, you have to use Gemini with it. A skill is portable; it is effectively a prompt or a set of prompts bundled together that you can move to any system that supports Skills. Anthropic made it a public good, open-sourced it, and said anyone can use the Skills format. There is a whole website called AgentSkills.io, which is the official definition of the Skills format.

They have a 32-page slide deck that explains what goes into making a good skill. Skills are supported by Claude Code, Claude Cowork, Google Antigravity, OpenAI, Codex, QwenCode, Open CoWork, and pretty much most agentic frameworks. Because they’re so useful, I wanted to take a step back. A skill fits into a specific place in the evolution of AI.

At Trust Insights, we’ve been working on this for quite some time. Originally derived from product-market fit, it’s evolved into its own thing with five levels. Level one is “Done by You.” That’s you and ChatGPT; you are the typing monkey and you’re just chatting—asking how to do things, the weather, where to eat, or if you should vacation in Montenegro. You’re doing all the work, but you’re learning prompting.

The second level is “Done with You.” These are Gems and GPTs. Effectively, you’re building a standard operating procedure into one of these things so that you have something like a 90s song lyrics generator as a standard operating procedure at the agent level.

Level three is project planning. On the podcast, we mentioned that the hottest programming language of 2026 is project management. Today, you have fully autonomous or mostly autonomous systems like OpenClaude and the managed agents built right into the Claude ecosystem. Katie, what do you suppose the input is for those managed agents?

Katie Robbert – 05:40

I would assume all of your skills, plugins, and project plans. As I’m saying it, I realize these each build on each other. In phase one, you start with the prompts. In phase two, you have prompts plus something. In phase three, you have plus and plus. You don’t just do something brand new in each of these levels; you build on what you did previously. You still have to master level one, level two, and level three. I would say I’m squarely in level three with a little bit of dabbling in level four.

Christopher Penn – 06:24

You’re going to call yourself wrong in just a moment because you already have the skills to do levels four and five. You may not know it, but you do. At level four, when you’re building something like OpenClaude, you’re providing a job description. A job description is a bundle of project plans plus skills, plus plugins. You say, “Agent, you’re my stockbroker. Go off and do it.” At that point, it’s like a full employee.

Level five is an organizational division with a strategic charter, where you have a swarm of agents with a supervisor agent running them. Someone like you, Katie, who has built multi-million dollar teams and managed people, has all the skills needed to create a strategic charter for an organizational unit. You could send that division off and say, “This is my marketing division. I’ve got an SEO agent, a PPC agent, an email agent, and a marketing manager agent.”

As that strategic charter, you provide all the job descriptions, skills, and plugins. The agent swarm goes off and comes back once a month to say, “Hey Katie, the virtual marketing department did all these things for you.”

Katie Robbert – 07:53

Having worked with humans, my guess is that if they only check in once a month and I look at the report, I’m going to say, “That’s not what I asked you to do. Why didn’t you do the things I wanted? You spent all this money.”

I can definitely see the potential in that. With that description, I’m squarely in level four at the moment. I can see your point about getting into level five, but as a human, I’m not totally comfortable being hands-off like that. That’s a “me” problem. With that understanding, where do skills fall?

Christopher Penn – 08:40

Skills fall in at level two. Skills are essentially SOPs. You can take any prompt, conversation, or raw materials from level one and bring them into a system to turn them into level two.

There are so many tools that do this right now. Claude Code and Claude Cowork both have plugin and skill builders. The ones that come in the package are good, but they’re not great. I want to talk through how I built one for myself that we put in the brand new Trust Insights Prompt Playbook. The existing builders operate on the presumption that what you’re putting in was good to begin with, which is not always the case.

Katie Robbert – 09:43

The out-of-the-box skills are generic. There’s no customization or tailoring toward what you, your team, or your organization does. John could use the legal or sales plugin, but he’s going to have to customize the instructions based on the terms we look for in contracts, what needs to be called out in an MSA, or our specific sales cycle.

Our clients don’t look the same as other clients, so we need to work with them and not force them into a certain process. For John, the sales and legal skills are good starts, but they won’t get him all the way there. It’s probably easier for most people to build their own skills for their specific documentation review or sales processes.

Christopher Penn – 11:07

The non-surprising answer is that if you want a good standard operating procedure, you should have a framework that governs that procedure.

Katie Robbert – 11:33

Is it the 5P Framework?

Christopher Penn – 11:35

Yes, that is the basis of the “prompt-to-skill” plugin we built. It takes your existing prompt, puts it through the five Ps, and you end up with a skill that gets auto-installed into your Claude environment.

You also get an understanding that is missing in most skill builders: an “evals” section that looks at whether or not the skill should fire. If you have similarly named skills, the triggering section—the YAML front matter of the skill—determines when the skill runs. If it’s badly written, your skill will fire at the most inopportune times and screw everything up.

The first thing you need to do is take the thing you want to turn into a skill and put it through the 5P Framework. Why are we doing this? What are we doing? Who’s going to do it and who’s impacted? How do you do it? Do you need supplementary tools like MCPs or web searches? And the critical part: the performance. Does this skill work as intended? A well-structured SOP covers all of these pieces.

Katie Robbert – 13:06

If you have well-structured SOPs, you’re already in good shape. The first time I was hired to manage Chris’s team, I asked to see the SOPs and they looked at me like I had eight heads. They didn’t have any. There is no real secret to building and growing teams; a lot of it is just good structure and SOPs. Having good SOPs will help you move more quickly through these five levels.

Christopher Penn – 14:06

Let’s look at “prompt-to-skill” in Claude Code. Every skill must have a skill.md file and a folder. The description tells the system when to fire. This one converts any existing prompt, workflow, or instructions into a structured Claude skill with scripts, references, and assets.

It doesn’t have to be just prompts inside a skill; you can have things like Python scripts. If you’re running those, be 100% sure you trust the source. In this flow, you run the 5P analysis: purpose, people, process, platform, and performance. It has to write down its analysis first because we never let an agent system do work only in memory—the moment you quit that session, it’s gone. It always has to take notes.

Then it determines the skill category, builds the name, writes the front matter with triggering capabilities, builds the body, and identifies resources. The fifth P is generating test cases. You have to test your skill. Inside the plugin is the eval. For example, if someone asks, “What’s the capital of France?” this skill should not fire. If it does, something’s gone wrong.

We baked this for ourselves in the Prompt Playbook so you don’t have to do it by yourself. No prompt exists in a vacuum. A critical question is: what are the upstream and downstream things around this prompt? A model like Opus is smart enough to infer that if John provides a Motown lyrics prompt, upstream it needs knowledge of good Motown songs, and downstream he’ll likely use a tool like Suno to generate a song. Therefore, it should also create a music composition prompt.

Katie Robbert – 18:19

What you’re describing are the basics of software development. A lot of people stayed away from generative AI because they aren’t developers. We call these things “skills” or “plugins,” but guess what? You’re developing software; you just don’t know it.

The other framework you need to lean on is the Software Development Life Cycle (SDLC). You start with requirements—that’s where the 5P Framework comes in. Then you have design, develop, test, and deploy. What Chris mentioned regarding performance, unit testing, and extraction is you testing the software you’ve built. If you aren’t testing as you go, you’re doing it wrong.

Systems can still hallucinate or make mistakes. Make sure there is validation at the end of every skill. You might build a skill today and pull it into a project three months later. If it’s a mismatch but there’s no testing built in, your results could go sideways. Be disciplined and strict about that.

Christopher Penn – 20:41

All the prompts in the Prompt Playbook can be dropped into the plugin to generate a skill. If you like a prompt from that week, you can just say, “Turn this into a skill.”

I’m going to take an old Gem for generating music for Suno. I’ll copy the instructions and say, “Let’s turn this Suno music generation prompt into a skill.” If you have instructions from ChatGPT or Google Gems laying around, you can get this to do a really good job.

In Claude Code, go with “plan mode” first. Use the Opus plan model for planning and Sonnet for implementation to save money. This fired up the “prompt-to-skill” because it correctly identified my intent.

Katie Robbert – 22:56

You must make sure your plan up front is as thorough as possible. If it’s building and installing it for you but it’s wrong—like accidentally programming it to delete your hard drive—that would be a very bad day.

John Wall – 24:09

It’s been funny because so many skills just show up in the interface. I know there’s magic involved, but I didn’t even realize there was “stock” stuff you can grab. I’ve been working with handcrafted tools and didn’t realize there was generic junk out there for the rest of the world.

Katie Robbert – 24:30

Chris and I are the magicians behind the scenes. By the time it gets to you, it’s been road-tested through five rounds of errors and lost stuff.

Streamyard now allows polls. They’re a little janky, but you can type your response into the chat. Our very scientific poll is asking who your favorite host is.

Christopher Penn – 25:44

It just finished its plan to convert the Suno V5 prompt guide into a skill. It did the 5P analysis. Apparently, John is overwhelmingly our favorite host.

Katie Robbert – 26:04

I can’t be mad at that because everybody loves John.

Christopher Penn – 26:17

It builds the “Prompt Architect” skill, evaluates the file structure, and identifies references like formulas and keyword dictionaries. It gathers musical intent, selects a formula based on a decision tree, and builds the style prompt, lyrics, and meta tags.

It’s important to figure out who’s going to use this—AI music producers, content creators, or hobbyists. If you infer those uses, it’s easier to define the expected performance.

Katie Robbert – 27:27

I appreciate that it writes out who is going to use it. We aren’t short on ideas, but we don’t always have a clear vision of who benefits from them.

Christopher Penn – 27:58

The source prompt is 4,000 words, but the skill budget is 500 lines. The extra details get moved into a reference folder. It’s writing this straight into my Claude folder, so I’ll have a music generator shortly.

Regarding the ideal folder structure: use the AgentSkills.io website. That contains the specs for how to lay out stuff. It is cross-platform compatible. Anthropic’s website also has a 33-page PDF guide to building Claude skills.

Now it’s building the keyword dictionaries and anti-patterns. “Anti-patterns” is a magic word AI loves. It’s the opposite of a design pattern—it’s the critical mistakes the system should not make. For me, using an em-dash is an anti-pattern, but that’s a rant for another show.

Katie Robbert – 30:53

If you’re using Claude Cowork or Claude Code and say you want to build a brand tone skill, it will ask what you should not say. That builds the anti-pattern database. If you want to write in a specific style, you have to tell it not to use certain phrases. Also, think of it in terms of scope—what is in scope and what is out of scope.

Christopher Penn – 32:22

If you’ve been following our advice, you should have a prompt library in Joplin, SharePoint, or elsewhere. You should not convert those prompts one at a time; that’s a terrible idea. Put the prompts as text files in a folder and tell a system like Claude Code to process every file and put it in an output folder.

Claude Code can directly interact with your hard drive. You have an input folder of your greatest hits and an output folder you define. You have a skill to make skills. Set it up before you go to bed, and the next morning your favorite prompts are ready to be deployed as skills.

If you are converting prompts one at a time, you are doing level one work with a level three tool. If you say, “Here is your project plan and the skills to convert these,” you are operating at the right level.

Katie Robbert – 35:00

You can use Claude CoWork to do nearly identical stuff if you aren’t comfortable in a coding environment. I have a whole set of folders on my desktop specifically for Claude. I had Claude organize and rename them. I can enact skills like the “Voice of the Customer” focus group to help name things. It’s weird, but it works.

Christopher Penn – 36:31

The one thing you can’t do in CoWork as easily as in Code is run command line tools. Google Workspace has a command line tool that can connect to Google Calendar, Drive, and Gmail. Your AI can use this. I could open Claude Code, use the GWS tool, and tell Claude to reach into our Google Keep, pull out our prompt library, and convert it all to skills. Microsoft has a similar CLI for OneDrive and SharePoint.

Katie Robbert – 38:08

A solid prompt library is table stakes. It is your foundation. Even if you build skills in Claude, do not get rid of your library. These tools change or implode. You need to keep your data somewhere. Your prompt library is your foundation, just like documented SOPs are the foundation for automation. If the automation breaks, you need to know where the process failed. John, you do sales, but we still have a sales playbook as a reference because you’re only one person.

John Wall – 39:37

We don’t have the capacity to remember everything. Since this is a protocol, can you grab these skills and throw them into other tools? How does that work in reality?

Christopher Penn – 40:03

I use QwenCode and Open CoWork, and the skills perform flawlessly because they follow the standard protocol. It is literally drag-and-drop.

The best practice for a company is to have a private GitHub repository as a skills marketplace. As you get better at prompting or new models come out, you adapt the skills. If you use a private GitHub repository, when you make changes, all the Claude instances in your company that have downloaded that skill auto-update. You don’t have to go computer by computer.

Katie Robbert – 41:42

We’ve struggled with that even as a small company. I changed a brand skill on my machine, but it didn’t update on Chris’s. That’s why centralizing them is vital.

Christopher Penn – 42:14

Ask your IT department if you have a company GitHub account and set up an internal marketplace. To learn more, subscribe to our Prompt Playbook at TrustInsights.ai/PromptPlaybook. We drop two new prompts every Tuesday and Friday. You can also find the 5P Framework at TrustInsights.ai/5P-Framework.

Our scientific poll shows John in the lead with 100% of the vote. Congratulations, John.

John Wall – 43:54

I may or may not have voted for myself. There are no rules against it. You can do all this work yourself, or be like me and hire people to do it for you. Talk to the fan-favorite host, and he will get you set up.

Christopher Penn – 44:42

Any final thoughts? The next evolution is building project plans composed of skills.

Katie Robbert – 44:57

If this feels overwhelming, start with your SOPs. Document how you complete any given task, put it in a central repository, and then you can start turning them into skills.

Christopher Penn – 45:19

One last thing: Claude Code can run “headless.” You could create a command to convert all prompts into skills and schedule it. On a monthly basis, a task could run on a computer to automatically convert and publish new prompts. This moves your company to level three very quickly.

Katie Robbert – 46:09

You can do the same in CoWork. I have two scheduled tasks running now. Take advantage of those options.

Christopher Penn – 46:57

That’s going to do it for this week. Thanks for tuning in. For more resources, check out the In-Ear Insights podcast and our weekly newsletter at TrustInsights.ai/Newsletter. Join our free Slack group at TrustInsights.ai/AnalyticsForMarketers. See you next time.