What is a "skill" in the AI context, and how do you build one for your workflow?

Question

Accepted Answer

A **skill** packages a reusable, on-demand workflow for an AI agent: **instructions** for a recurring task (when to do it and how), plus **optional scripts or resources** the agent loads only when the task is relevant. It is the difference between re-explaining a procedure every time and teaching it once.

## What a skill is

```text
SKILL = a self-contained folder the agent can pull in on demand:
  SKILL.md   → metadata + instructions (the TRIGGER + the STEPS)
  scripts/   → optional helper scripts the agent runs (e.g. a linter, a generator)
  resources/ → optional templates, examples, reference files
→ loaded ONLY when relevant → keeps the base context small, the workflow repeatable
```

The payoff is **progressive disclosure**: the agent knows the skill *exists* from a short description, and pulls in the full instructions and scripts only when a matching task shows up — so dozens of skills don't bloat every conversation.

## Building one

A skill has three parts: the **trigger** (when), the **steps** (how), and any **helper scripts** (tools). Keep it focused and composable — one job done well, not a kitchen sink.

```markdown
---
name: changelog-entry
description: Add a CHANGELOG entry. Use when finishing a user-facing change.
---
# Add a changelog entry

## Steps
1. Read the latest commits since the last release tag.
2. Categorize each as Added / Changed / Fixed.
3. Prepend a dated section to CHANGELOG.md under [Unreleased].
4. Run `scripts/validate.sh` to check the format.
```

The `description` is the **trigger** — it tells the agent when to reach for this skill. The numbered list is the **steps**, and `validate.sh` is a **helper script** that enforces correctness so the agent doesn't have to reason about format from scratch. Because the skill is small and single-purpose, it composes cleanly with others (e.g. a separate "release" skill can call it).

## Why it matters

Skills turn an agent from a general assistant you re-brief constantly into one that knows *your* recurring procedures. By packaging the trigger, steps, and scripts in a focused, composable unit that loads only on demand, you encode institutional knowledge once and reuse it reliably — getting consistent results on repetitive tasks without re-explaining them, and without paying the context cost of every workflow on every request.