Skills: Your AI Now Knows When to Act on Its Own

A skill looks like a prompt. It isn't.

A prompt you write, send, and the AI forgets the moment the conversation ends. A skill stays saved, and the one who decides when to use it isn't you, it's the AI itself. You can have ten skills saved at once, each one ready for a different task. The AI reads only the short description of each before starting, and the moment a task matches one of them, it loads the full content on its own, without you having to point at which one to use.

It's that difference, a prompt you repeat versus knowledge the AI recognizes on its own, that got OpenAI, Google, Microsoft, and Anthropic to launch, almost at the same time, a similar answer to the same problem, each one under a different name.

OpenAI called it custom GPTs: you build a persona with fixed instructions and reference files, save it, reuse it. Google did the equivalent with Gems in Gemini. Microsoft went a step further with Copilot agents, which don't just respond with fixed instructions, they also trigger actions inside Microsoft 365, like scheduling a meeting on their own. And Anthropic launched Claude Skills, which solve the problem in a structurally different way: not a saved persona, but a folder of instructions, with scripts and supporting material, that the AI loads on its own only when the task calls for it.

So far, this looks like four companies copying the same marketing idea, each slapping its own name on it to look like exclusive innovation. It isn't.

The proof this isn't a positioning coincidence

GitHub Copilot, Microsoft's own product, now reads skill folders in the same open format Anthropic published, Agent Skills, at agentskills.io. A direct competitor adopting the other's standard isn't a marketing coincidence, it's a sign the problem is real.

And here's an opinion: custom GPTs and Gems solved half the problem. They saved typing, but they were still text the AI re-reads from scratch every conversation. Skills go further, they carry procedure and scripts, and only step in when the task calls for it. It's the difference between handing someone a manual to reread every time and having a trained employee who already knows the right moment to apply each rule.

Now picture a support lead who always knew, by heart, how to classify ticket urgency, but never had anywhere to store that so the AI could use it on its own. His skill would be a folder named something like "support-triage," with a file inside that looks like this:

---
name: support-ticket-triage
description: Use when the user pastes a ticket, email, or customer message asking to classify urgency and suggest an initial response.
---

# Support Ticket Triage

## Steps

1. Classify urgency into three levels:
   - **Critical:** system down, customer locked out.
   - **High:** broken feature, no workaround.
   - **Normal:** question or adjustment request.

2. Ask whether the issue has appeared in another recent ticket and request the history if there is one.

3. Suggest a short initial response, acknowledging the issue and stating the expected response timeline for each level:
   - **Critical:** within 1 hour.
   - **High:** within 4 hours.
   - **Normal:** within 24 hours.

4. Never promise a technical fix that the team hasn't confirmed yet. Only confirm the response timeline, not the resolution timeline.

## Output Format

**Urgency:** [level]

**Suggested response:** [ready-to-copy text]

**Internal note:** [something the team needs to know before responding]

What's in the file is only the tip of the iceberg. After loading these instructions, the AI also checks whether the folder contains any scripts to run (such as an automatic deadline calculation or a system check) and whether there are any additional reference materials associated with it (like a historical spreadsheet or a more detailed guide). Those resources are only brought into the conversation if the task actually requires them. Instead of loading everything at once, it loads information layer by layer, based on what the task demands.

From the next Monday on, he just pastes the ticket. The AI recognizes the task as triage, loads that folder on its own, and applies the standard without him having to rewrite anything.

That doesn't mean everything that repeats twice deserves to become a skill. Most of what people ask AI for is still a one-off request, not worth the effort of packaging. The skill that matters now isn't just writing a good prompt at the right moment. It's recognizing, inside your own routine, what repeats enough to justify becoming stored knowledge, and what's better left as a one-off request.

Whoever confuses the two ends up wasting time packaging what should've just been asked directly, or keeps re-explaining, every Monday, something that should've been solved months ago.