Secret sauce

When I started a new role at Google earlier this year, I ran into a data problem. I’ve been at the company for 15 years, so I wasn’t exactly starting from zero. I knew how our large datasets worked, I knew our internal systems, and the documentation for my new team was genuinely pretty good. The problem was understanding the team’s massive collection of data. The dataset had more tables than I could hold in my head at once, and the joins only made sense after you’d studied hundreds of pages of docs. I tried writing it down and drawing graphs to visualize, but nothing clicked the way I needed it to.

Eventually I started mapping each table into a markdown file. Every field got a name, a type, and a few rows of example data. Once that was solid I added in how the tables connected to each other in a join section. I then handed the file to an AI and asked it a question. “How many tickets did we have open last month, grouped by type and priority?"

Gemini spit out a full SQL script in about three seconds.

Skills

Most people open an AI website, see the text box, and start typing. The assumption is that the magic is in the question. That’s why we see “Prompt Engineer” on employment boards. People assume that if you ask your question in exactly the right way, you’ll get the perfect answer. So they rephrase, retry, and iterate until something fits. Sometimes it works, but most of the time it falls short.

The part nobody considers is what the AI doesn’t know. It doesn’t know who you are, what you’re working on, or why you’re asking. Every session starts from zero and a skills file is how you can fix that.

A skill is just a plain text document, written in whatever format makes sense for the content. Mine are markdown files, and I keep them in my note-taking app Mark so they’re on a server, available on every device, and easy to update. When I want to use one I download it and add it to the conversation as context.

The best way I’ve found to explain what a skill actually does is my oncology file. I keep extensive notes on my health, and maintain a running journal of everything related to my treatment:

• Full prescription history, past and current
• Doctor details
• Summary of every blood test and scan
• Things I need to remember to ask about
• Blood pressure and oxygen readings over time
• Additional notes

When my doctor wants to change a medication, I open that file, hand it to an AI, and ask whether the new drug has side effects that conflict with anything I’m already taking. I get a real answer based on my actual situation, not a generic AI disclaimer to “consult a physician”. It lets me show up to appointments as an informed participant instead of someone just hoping the people in the room know what they’re doing.

A skill is a journal you wrote for yourself that an AI can also read.

For people who work in tech, I use the doc coauthor skill as a great example. I use it heavily at work and constantly adapt it to my own needs. After you give it to AI, it starts by asking you for a brain dump of everything relevant to what you’re writing. That includes notes, artifacts, random observations, and half-assed ideas. It takes all of it, helps you figure out what kind of document you want, and walks you through building it section by section. When the draft is done, it reads the whole thing from the audience’s perspective, finds the gaps, and asks follow-up questions you might get in a meeting.

Examples

I intentionally mention both the SQL doc and the oncology journal as they’re perfect examples of AI skills. One is a technical reference built for a specific dataset, while the other is a personal health journal. The fact that both work as skills demonstrates the point I’m trying to make. It gives the AI context it wouldn’t otherwise have.

A third skill I use constantly is a humanizer. AI writing has very specific tells (e.g. emojis, emdashes, colons, lists). All of it comes from being trained on an aggregated average of the world’s writing. The training sources affect the words it chooses, the way it puts together sentences, and the overall mechanical rhythm. They’re usually written fairly well, but it’s just not human or anything close to the way we speak. A humanizer skill is a document that defines what your writing actually sounds like and removes the AI patterns. Mine has taken a long time to build and I’m constantly making adjustments. I fed it examples of my own writing, worked with the AI to identify what made it distinct, and refine it as I found new patterns to adjust. The result isn’t perfect, but it means I spend consideraly less time editing and more time focused on what I’m actually trying to say.

That’s the better way to think about what skills do in general. For example, most people building a presentation will spend the majority of their time on formatting. The slides get a lot of attention on how it looks, but the message gets buried. I spend way too much time tweaking text boxes and images so they line up and not enough time on content. A document creation skill flips that. The structure and formatting get handled. You focus on the content, which is really the only part that matters.

A skill can be a database schema, a medical journal, a voice profile, a list of your recurring meeting attendees, or a summary of a project you’ve been on for two years. Whatever context you find yourself re-explaining to the AI every time you start a new session, write it down and use it as context. In fact, you can merge all your existing notes into one skill file and let AI run with it.

Use anywhere

A lot of people pick an AI and treat it like a platform commitment. They learn its quirks, build a workflow around it, and assume moving to a different one means starting over. Skills are one of the reasons that’s not really true.

At work I use Gemini because that’s what the company pushes. At home I use Claude for anything involving code and ChatGPT for quick questions - the kind where I need to know which actor is in the movie I’m watching or what to eat around a specific medication. The same skill files work across all of them. Almost every AI interface has an upload button. You attach the file and it instantly becomes context for the discussion. The conversation will feel different depending on which model you’re talking to, because each AI has different training data and different restrictions, but the skill will work across most all of them.

The same idea extends into more advanced tools. Cursor, Windsurf, Lovable, and the growing category of vibe coding tools that let you describe an app in plain language and watch it get built can all work with skills. If you need consistency across sessions in any of these tools, a skill is how you get it.

Mine

The best way to understand what a skill can do is to see a few real ones.

HealthCoach: This is my health journal mentioned earlier. The file opens with a patient profile, then moves through diagnosis, treatment regimen, medications, known side effects, symptoms, and next steps. Every section is written plainly so the AI can read it without interpretation. When I have a question about a medication change or a new symptom, I attach this file and prompt. The AI has everything it needs to give me a useful answer instead of a generic one. The skill gets updated after every appointment, message, and test result. (Download example health skill)

SQL skills: These are built around the internal datasets I work with at Google. It documents the key tables, what each field contains, an example value for each one, and the relationships between tables. When a new analyst joins the team, understanding how these tables connect can take weeks. This skill compresses that into a single file. I attach it, describe what I want to know, and get a working query. The file gets updated whenever the schema changes or a new use case needs documenting. (Download example SQL skill)

Smart Caveman: This is the skill I use for writing and coding assistance. It’s a large file, but has a lot of unique jobs. The first is compression to preserve token usage. It strips the AI’s default output of filler phrases, sycophantic openers, and any fluff associated with AI chat. Things like “Great question!” and “It is important to note that” get cut automatically so you’re not charged usage fees for nonsense. The second job is voice matching. I trained it on existing blog posts and writing samples until it could produce a first draft that sounds like me rather than like an AI averaging across the internet. It also runs a workflow that collects a brain dump at the start, works through document / email / blog post creation section by section, and reviews the finished draft from the audiences’s perspective to catch any logical gaps. (Download example caveman-voice skill)

Diet skill: This is one I’m still actively working on. It’s a meal estimation skill that tracks calories and protein throughout the day. It can expand into other metrics too, like macros, water intake, and exercise. Right now it does exactly what I need. The output is minimal, and it tells me how confident the estimates are. If confidence is low, it asks follow-up questions to give me a better result. I use it multiple times a day and have it handy on my bookmark dashboard “Dash”. (Download example meal tracking skill)

Your turn

The fastest way to build a skill is to just start a conversation. Open any AI and say something like “I want to create a new skill. Here’s my idea.” Then describe it. The AI will ask questions, help you figure out what belongs in the file, and work through it with you.

Skills are plain text files, usually written in markdown. Markdown is a simple formatting system where a line starting with # becomes a heading, **this** becomes bold, and a dash at the start of a line becomes a bullet point. You don’t need to know it well to write a skill. Most AIs will accept a plain text file with no formatting at all and still understand it just fine. Markdown just makes it easier to organize. Keeping notes in Markdown is incredibly helpful and is how my note-taking app Mark manages content.

If you want a more structured starting point, Anthropic published a skill creator tool and a full skills directory on GitHub with a lot of examples. There are also marketplaces like skillsllm.com and skillsmp.com that are worth checking out. Both of these are built around open-source skills for Claude, Gemini, and ChatGPT.

The best opportunity for a skill is when you open a new session and find yourself pasting the same paragraph of context every time. That paragraph can become a reusable template you save in a doc. I’ve built them for many scenarios. Each one exists because I got tired of re-explaining the same thing.

When you sit down to write one, three things matter. The first is a clear name and a one or two sentence description of when to use it. The second is the context itself. This is whatever the AI needs to know, written the way you’d explain it to a knowledgeable colleague on their first day. The third, and optional, piece is behavioral instructions. This is where you tell the AI how to respond, not just what to know. My Smart Caveman skill works this way. It doesn’t just give the AI information, it tells the AI which mode to use, what patterns to strip, and how to match a specific writing style.

Start small. A single-purpose skill that does one thing well beats a massive file that tries to do everything. You can always add to it later. Come to think of it, I should take my own advice here. Smart-Caveman could be broken into multiple docs to save on tokens. Skills can reference other skills.

Next up

Everything covered here should get you off to a great start. The next, and more advanced layer is automation. Tools like Claude Code can take a skill beyond context and into action, running commands, managing files, and executing tasks on your behalf without you prompting each step. The skills in the further reading section below will push you into that journey.

Start with any of the examples in the “Ones I’ve Built” section above, or download one from below. Fill in your details, play with it for a week, and see if you dig it. The path from a simple markdown file to a fully automated workflow isn’t necessarily a long one, but it always starts in the same place.

Links & Sources

Here are some skill resources that you should bookmark.

• Anthropic skills directory — The official collection from Anthropic. A good starting point for ready-to-use skills and structure examples.
• Skill creator — Anthropic’s tool for walking through building a new skill from scratch.
• Doc coauthor — The Anthropic skill I use for drafting documents. Collects a brain dump, proposes structure, works section by section, and reviews the finished draft from the reader’s perspective.
• Humanizer — Strips AI writing patterns from text. A good base to build a personalized version from.
• Get Shit Done — Addresses context rot, the quality drop that happens as a long session fills the AI’s context window.
• Superpowers — A full software development methodology built on composable skills for coding agents.
• The Agency — A collection of specialized agents, each with its own personality, process, and purpose.
• skillsllm.com and skillsmp.com — Community marketplaces for open-source skills across Claude, Codex, and ChatGPT.

Feel free to reach out if you have any questions or want to share your experience with AI skills.