How do AI skills work?

AI skills work by giving an agent a reusable workflow it can load for a matching task. The agent first sees a short description of the skill. When the task fits, it reads the full SKILL.md instructions and uses any referenced scripts, documentation, examples, or templates. The skill shapes the process; the model still performs the reasoning and writing.

That distinction explains both why skills are useful and why they are not magic.

Step 1: the agent discovers the skill

A skill has to be installed somewhere the agent knows to inspect. That might be a personal skills directory, a folder inside one repository, or a plugin-specific location.

During discovery, the agent does not need to read every file from every installed skill. The open Agent Skills specification uses progressive disclosure: the agent can start with the skill's name and description, then load deeper instructions only when needed.

This keeps a collection of skills available without placing every rubric, script, and example into the conversation up front.

Installation paths depend on the product. Use the dedicated setup guides for Codex or Claude Code rather than assuming one folder works everywhere.

Step 2: the description routes the task

The description in SKILL.md tells the agent what the skill does and when it should be used.

Compare these two descriptions:

Helps with websites.

Audits public landing pages for crawlability, search intent, metadata,
internal links, unsupported claims, and conversion gaps.

The first can match too much or nothing useful. The second names the task and the checks that distinguish it from general writing or web development.

Good descriptions also set boundaries. An SEO audit skill might explicitly exclude backlink outreach or keyword-volume research. A support-reply skill might draft messages but never send them.

This is why the metadata is part of the workflow, not a marketing summary pasted above it.

Step 3: the skill is activated

Activation can happen in two ways.

You can invoke a skill explicitly by naming it in your request. That is the clearest method when testing a new package or running a sensitive workflow.

An agent can also select a skill because your normal-language request matches its description. You might ask, "Review this release against our App Store submission checklist," and the agent can load the relevant release skill without you naming the file.

Automatic selection is convenient, but it needs testing. A description that is too broad can interrupt unrelated work. One that is too narrow can miss requests that should use it.

Step 4: the agent reads the working method

After activation, the agent loads the Markdown body of SKILL.md. This is where the author defines the actual job.

Useful instructions often specify:

• Inputs to inspect before starting.
• Project or domain context that controls the answer.
• Steps that must happen in order.
• Decisions the agent can make on its own.
• Situations that require clarification or approval.
• Output format and completion checks.

Consider a content-refresh skill. It might require the agent to inspect the target query, current page, Search Console export, competing intent, internal links, and conversion goal before suggesting edits. It might tell the agent to preserve factual claims, flag missing evidence, and return a change plan before rewriting.

The same model given "make this article better" has much more room to guess.

Step 5: supporting files are loaded when needed

SKILL.md can point to files stored beside it.

A workflow may include:

• A reference file with a 40-point audit rubric.
• A template for the final report.
• Two annotated examples that show the quality bar.
• A script that extracts metadata or checks a folder.
• A lookup table used during classification.

The instructions should say when each file matters. Dumping ten references into a package without navigation moves the confusion from the chat into the folder.

Supporting files are one of the clearest differences between a skill and a saved prompt. They let the workflow carry detailed material without turning the entrypoint into one enormous block of text.

Step 6: tools and scripts do the mechanical work

Some skills are instruction-only. Others ask the agent to run scripts or use connected tools.

A script can collect page titles, validate a JSON file, resize screenshots, or check whether required sections exist. An external tool might retrieve analytics, open a pull request, or read support tickets.

The skill should state those dependencies and permissions before the action happens. Instructions cannot provide credentials or system access by themselves. The difference between workflow and access is covered in AI skills vs MCP servers.

Treat bundled scripts like any other code dependency. Read them, check their inputs and write paths, and test them on low-risk data first.

Step 7: the result is checked

A good skill explains what "done" means.

That might be:

• Every finding cites a source line or page element.
• A generated file passes a schema validator.
• The report separates evidence from inference.
• Tests and lint pass after a code change.
• A draft remains unsent until a person approves it.

Validation does not make model output deterministic. It catches missing steps and packaging errors, and it gives the agent a visible standard to work toward.

Skills do not train or replace the model

Installing a skill does not retrain the model or permanently add new intelligence. It supplies context and instructions when the workflow is activated.

A stronger model may reason better about a difficult input. A better skill may make the same model more consistent on a repeated job. Both can matter, but they solve different problems.

Start by asking what failed. If the agent cannot do the underlying reasoning on a focused example, the model or tool may be the limit. If it can do the pieces but keeps missing your process, a skill is a sensible fix.

For the broader category definition, read what AI skills are. Then test any installed skill with one direct invocation, one natural-language match, and one nearby request that should not trigger it. That small test tells you more than the folder name does.