Everyone's talking about them, but a lot of the explanations floating around blur them together with something else entirely, usually a Claude Project with a good prompt attached. Here's the real difference, how to get an actual Skill working today, and how the same idea shows up in Gemini and ChatGPT.
You've probably seen ads for "Quarterly Goal Recalibrators" and "Timeblock Architects" and thought: what exactly am I looking at? Let's cut through it, correctly this time.
A Skill is a small package Claude can pick up and use. Inside it is a file called SKILL.md holding instructions for one specific job, sometimes with extra scripts or reference files bundled alongside it. You build it once, and it works everywhere you use Claude: on claude.ai, inside Claude Code, and through the API.
The part that sets a Skill apart from anything you'd type into a normal chat is that it loads on its own. You don't have to remember to open the right folder or paste in the right instructions first. When you're doing the kind of task a Skill covers, Claude notices and pulls it in. When you're not, it stays out of the way and doesn't clutter up the conversation.
The short version: A Skill is a reusable set of instructions Claude loads automatically, only when it's relevant, and that follows you across every way you use Claude. Skills work across claude.ai, Claude Code, and the API. Anthropic has been widening which plans get access, so check Settings if you don't see them yet.
If someone sold you a "Claude skill" for real estate and what you actually got was a link to a shared Project, you didn't get scammed exactly, you just got the wrong label on the box. Here's the real difference.
A Project is a workspace with a system prompt and some files attached. It only works when you open a chat inside that specific Project. Start a regular chat instead, and none of it applies. You have to remember where you put it.
A Skill activates on its own, in any chat, in Claude Code, wherever you're working. Claude recognizes the task and pulls in the right instructions without you having to go find the right workspace first.
None of this means the Project method is bad. It's still one of the easiest ways to get a consistent, specialized version of Claude working for you, and it's worth building even if you never touch a real Skill. We walk through it below, correctly labeled this time.
None of these require writing code. Pick whichever matches how much time you want to spend on it.
In Claude.ai, go to Customize, then Skills, then Browse skills. You'll find Anthropic's own document skills for Excel, Word, PowerPoint, and PDFs, plus partner-built ones from companies like Notion, Figma, and Atlassian. Toggle one on and it's ready the next time it's relevant. No setup beyond that.
Start a chat and ask Claude to help you build a skill for something you do repeatedly. Claude interviews you about how you actually do the task, then writes the SKILL.md file for you. No code required for anything reasonably simple.
Claude's Cowork mode added a screen recording option: you do the task once on your own computer while it watches, and it turns what you did into a repeatable skill. You never type out a single instruction. It's the newest option here and still rolling out, so you may not see it depending on your plan.
This is the Project method from the comparison above, not a literal Skill upload. There's nothing to toggle on and nothing to record. It still gets you a specialized, consistent Claude today, and it's the easiest on-ramp if you'd rather not touch a new setting yet. We'll build a Listing Description Writer as the example.
In Claude.ai, look for "Projects" in the left sidebar. Click "New Project" and give it a name. Something clear and functional works best: "Listing Description Writer" beats "My Real Estate Tool."
You'll land inside the Project workspace. This is where you'll write your instructions and upload your context before starting any conversations.
Click "Edit Project Instructions" (sometimes labeled "Set Instructions" depending on your interface). This is where you tell Claude what this Project is and how it should behave every single time.
Use the RCTOA framework to structure it: Role, Context, Task, Output, Adjusters. Here's a starting point for the Listing Description Writer:
Still inside the Project (not inside a conversation), look for the option to add files or documents. Upload anything Claude should always know about when using this skill.
For a Listing Description Writer, useful uploads might include:
No documents yet? Start without them. You can always add context over time as you figure out what Claude keeps getting wrong or asking you to clarify.
Now open a new conversation inside that Project. Give Claude your first property details, the way you naturally would. See what comes back. This is your first test run.
If the output isn't right, don't just keep prompting inside the conversation. Go back and update your system prompt or add a context document. Fix the source, not the symptom. That's what makes a Project actually hold up over time.
The first version of your Project is not the final version. Every time Claude misses the mark, ask yourself whether the issue is in your system prompt or your context documents. Update accordingly. A well-maintained Project gets better with use because you get better at defining what you actually need.
Some agents keep a running notes document that they update and re-upload every few months: new neighborhoods they've added, updated pricing context, style adjustments. That's how a Project stays sharp. And once you know the instructions actually work, turning that same system prompt into a real Skill with skill-creator takes about ten more minutes, and it'll follow you outside that one Project too.
These are starter instructions for three common real estate tasks. Use them as a Project system prompt with the method above, or hand the same text to skill-creator to turn it into a real Skill. Click any card to see the prompt, copy it, and put it to work.
Feed it your numbers each quarter. Get back a clear-eyed read on where you stand, what your run rate needs to be, and what to adjust if you're behind.
Give it the situation and the relationship. Get back a follow-up that sounds like you wrote it, not like it came out of a template folder.
Give it your sphere breakdown and your weekly availability. Get back a concrete 30-day contact plan with specific outreach angles for each segment.
If your brokerage or your team is on a different assistant, the same concepts exist, just under different names and with different limits.
A Gem is Google's version of the Project method: a saved name, instructions, and knowledge files that apply every time you open that Gem. Separately, Gemini in Chrome added a feature actually called Skills this year: you save a prompt straight out of your chat history and reuse it with a slash command on any webpage. Useful for a repeated one-off task, but it's a saved prompt tied to the browser, not a package that follows you into every Google product.
A Custom GPT is a shareable, pre-configured bot built with a system prompt, files, and tools, similar to a Claude Project but built to hand off to other people. ChatGPT Projects work more like a persistent folder for related chats and files. Apps in ChatGPT are third-party integrations, think Zillow or Spotify, that surface when they're relevant, which is closer in spirit to a Skill activating on its own, except you're using someone else's app rather than building your own reusable workflow.
Bottom line: every major assistant has some version of "save a setup and reuse it." What's still harder to find outside Claude is a personal Skill you build once that follows you into every surface that assistant offers, not just one workspace or one browser tab.
The difference between something you actually use every day, whether it's a real Skill or a well-built Project, and one that collects dust usually comes down to the same four things.
Specificity of context. A skill that knows your market, your price range, your typical client, and your communication style will always outperform a generic one. "Real estate agent" is not enough. "Buyer specialist in Hanover County Virginia working $350K-$600K with first-time buyers and relocators" is a skill that knows where it lives.
Clear constraints. The best system prompts define what Claude should never do as clearly as what it should always do. "Never use the word stunning" is more useful than "write good descriptions." Constraints make outputs consistent.
A defined output format. If you want a short MLS field AND a full remarks section, say that. If you want three options to choose from, say that. If you want bullet points, say that. Claude defaults to whatever seems reasonable. Tell it what you actually want.
Maintenance habits. A skill built in January with January's market context is stale by June. The agents getting the most out of this treat their context documents like a living file: update the numbers, add new neighborhoods, adjust the tone guidance when you notice drift. Ten minutes of maintenance every quarter keeps a skill sharp.
The RCTOA connection: If you've used the WiredToClose prompt library, you already know the RCTOA framework. Role, Context, Task, Output, Adjusters. Writing a skill's instructions is just RCTOA at a higher level. You're not writing a prompt for one task, you're writing a persistent setup for a category of tasks. The framework holds.
Not ready to build a full skill yet? The prompt library has 50 tested, framework-quality prompts you can use right now. Each one is a solid starting point for a future skill.
Browse the Prompt Library