How do you feel when someone says, “I need you to get this proposal ready for tomorrow”?
If you've ever prepared a quote or a sales presentation, you know that writing the content is only part of the job. You also have to check the information, follow your company's visual identity, choose the right format, and make sure nothing important is missing.
You can ask an AI assistant for help. You might even have a carefully written prompt saved somewhere. But whenever the document type, recipient, or requirements change, you end up adjusting that prompt, finding examples, and explaining how you want things done all over again.
What if you could teach the assistant that process just once?
Imagine simply saying, “Prepare the proposal,” and having your assistant already know which template to use, how to write it, and what information it needs before getting started. Not because it can guess what you want, but because you previously taught it the process.
That's the idea behind skills for AI agents.
Teach it a process, not just ask for an output
A skill is a reusable set of instructions and resources that an agent can use to perform a class of tasks according to a defined process.
Unlike a prompt that you copy and manually adapt, a skill lets you organize instructions alongside reference files, examples, and tools. In compatible environments, the agent can recognize when a skill is relevant and load the resources it needs without keeping all of them permanently in the conversation.
Think of a company that regularly prepares proposals, presentations, and follow-up emails. Each communication has a different purpose, but they share standards: tone of voice, visual identity, commercial terms, and presentation formats.
A skill can bring those rules and templates together so that, when someone requests a proposal, the assistant knows how to prepare it. If the price or scope is missing, it should ask before proceeding. If it needs a corporate HTML template, it can read that file instead of trying to recreate the design from scratch every time.
The difference isn't that the AI becomes smarter. It's that we stop relying on improvised instructions and start working with a known process.
How to create your first skill
You don't need to train a model or build an application. With the open Agent Skills format, a skill can begin as a folder containing a Markdown file named SKILL.md.
Let's create a skill for a company's communications. Not every message serves the same audience or purpose: an internal message needs a different format from a client proposal, and a newsletter has its own structure.
We can organize the resources like this:
communications/ ├── SKILL.md └── assets/ ├── internal.html ├── client.html └── newsletter.htmlThe main file explains when to use the skill, how to identify the communication type, and which rules to follow. The HTML templates preserve the visual format of each document.
Here's a simplified SKILL.md:
--- name: communications description: >- Prepares internal communications, client messages, and newsletters using company templates. --- # Company communications ## Template selection Identify the communication type: - INTERNAL: communication between company teams. Use assets/internal.html - CLIENT: proposals, quotes, and commercial follow-ups. Use assets/client.html - NEWSLETTER: informational messages for subscribers. Use assets/newsletter.html If the type is unclear, ask before selecting a template. ## Process 1. Identify the type and recipient. 2. Select the appropriate template. 3. Review the goal and available details. 4. Request any missing required information. 5. Write in a tone suited to the recipient. 6. Fill in the selected template. 7. Deliver the result for review. ## Rules - Never invent prices or terms. - Preserve the template's design. - Match the tone to the audience. - Never send without authorization.Now we have something more useful than a reusable template: a small process that can choose different paths depending on the request.
If we request an internal announcement, the agent uses the team format. If we ask for a quote, it selects the client template. And if we need a newsletter, it uses the corresponding structure.
This selection can be handled through instructions, as in our example, or through code when strict rules are required. Classification performed by an AI model is not deterministic validation.
We can extend the skill with examples of previous communications, writing guidelines, or scripts that check required fields. It can also rely on external tools or other skills, as long as the environment supports them.
Each resource has a clear responsibility. If the visual identity changes, we update the templates. If we introduce a new communication type, we extend the selection rules. We don't have to rebuild everything each time.
Why not just save a prompt?
For a one-off task, that may be enough. Skills don't replace prompts, and they don't guarantee perfect results.
Their advantage becomes clearer when we need to repeat processes and maintain consistency. Instead of pasting long instructions into every conversation, we can keep them organized and let the agent consult only the relevant resources.
This mechanism, known as progressive disclosure, helps avoid unnecessary instructions and makes maintenance easier. It does not, however, guarantee an isolated context or eliminate mistakes and malicious instructions from external content.
It also doesn't mean every response will be identical. Generated content will still vary, but we can reduce that variability by using templates, clearly defined inputs, and programmatic checks.
If a document's design must remain exactly the same, it's better to preserve its HTML structure and let the AI change only the designated fields. There's no need to delegate decisions that can be handled deterministically.
From one skill to a workflow
Once we've learned to organize a task, we can start combining capabilities.
Our communications skill could retrieve information about a client through an authorized tool, prepare a proposal, use another capability to generate a PDF, and finally check that the document contains the required information.
Each part can evolve independently, without turning the agent into a system that must figure out absolutely everything on its own.
But before getting there, start small and test it. Ask for an internal communication, then a quote with missing information, and finally a newsletter. Check whether it selects the right templates, and refine the instructions where needed.
A good skill isn't the one with the most instructions. It's the one that helps complete a task clearly, consistently, and in a way that can be checked.
The real value is in the work we stop repeating
Turning a process into a skill means taking the time to understand how we work, which decisions we make, and what standards we use to judge an outcome.
That knowledge is often scattered across documents, examples, habits, and people. Organizing it can be valuable even before we automate anything.
When we make it available to an agent, we can begin to change the way we work. Instead of explaining every step of a familiar task, we can focus on the result, review what matters, and spend our time on decisions that genuinely require our attention.
Perhaps that's where the real potential of skills lies: not in making AI do absolutely anything, but in teaching it to perform the work we know and repeat every day.
Further reading: Agent Skills — Open specification · Anthropic — Agent Skills