SKILL.md for new users · Field guide
No. 03
2026
Part 03 · The Decision

Skill or agent? Pick the right one.

In this part

How a SKILL.md differs from a regular agent, the one question that picks between them, a decision matrix for real cases, and a recap to take into your first authoring session.

A skill is knowledge the agent loads. An agent is a worker you summon. Most teams confuse them, and most over-build agents when a skill would have done the work.

Different shape, different job.

A regular agent is a configured persona with its own tool budget, conversation context, and lifecycle. A skill is a folder of expertise that any agent can pick up and use. They're not competitors. They're different layers.

Layer · Knowledge

SKILL.md

A folder. A markdown file with YAML frontmatter. The agent loads it on demand when your request matches its description.

  • Loaded by description match — lazy, not always on.
  • Has no tools of its own; it tells the host agent what to do.
  • Reusable across many agents and many conversations.
  • Cheap to write, cheap to delete, cheap to version.
  • Lives next to the code it serves.
Layer · Worker

Agent (.agent.md)

A configured role with its own identity, system prompt, tool allow-list, and (sometimes) its own model. You summon it to do a task; it runs and reports back.

  • Always present when invoked — owns its own conversation.
  • Has tools and constraints declared upfront.
  • Can be a subagent: the main agent delegates work to it.
  • Heavier to author; pays off when the role is reused.
  • Often consumes skills to do its job.

The headline distinction: agents do things; skills tell whoever is doing the work how. A good agent is small and consumes many skills. A good skill is small and gets consumed by many agents.

Ask: do I need a worker, or just expertise?

If the answer involves a new persona, new tools, or a delegated subtask the main agent shouldn't handle directly — that's an agent. If the answer is "the main agent just needs to know how to do this kind of work" — that's a skill.

A few sharper tests:

  • Does it need a constrained tool list different from the host agent's? → Agent.
  • Does it run in its own conversation and report back a single result? → Agent (specifically a subagent).
  • Is it just how to do a task the host agent is already capable of? → Skill.
  • Does it bundle templates, examples, or scripts that get loaded only when needed? → Skill.
  • Do you find yourself writing "act as a…" — meaning you want a different voice or scope? → Agent.
  • Are you writing "when X, do Y, then Z" — a procedure for the current agent? → Skill.

Real cases, picked.

Use this as a sanity check before you scaffold anything. Most "I need an agent" requests turn out to be skills in disguise.

You want…PickWhy
The agent to follow our internal release-notes format whenever it drafts release copy. Skill It's a workflow the host agent runs. No new tools, no new persona. Lives in release-notes/SKILL.md.
A read-only code explorer that searches the repo and returns a summary so the main chat doesn't get cluttered. Agent A subagent. Its own conversation, its own tool scope, returns one report. The main agent stays focused.
Convert a mood-board into HTML tokens using a procedure the team has refined. Skill Bundled knowledge + a procedure + sometimes a starter file. Exactly what skills are for.
A "code reviewer" that reads diffs against project standards and writes severity-graded findings. Both The reviewer is an agent. The standards it checks against are skills the agent loads.
Tell the agent never to use --no-verify on git pushes. Neither — instructions Always-on rule that binds to certain files. That's an .instructions.md, not a skill or agent.
A "PRD builder" persona that interviews you and produces a finished PRD. Agent It needs a persona, a conversational lifecycle, and a specific final artifact. Agent territory.

The skills you'll write first.

If you're new and looking for a place to start, almost every team's first three skills fall into one of these shapes.

Archetype 01

The style guide.

Your brand voice, your error-message tone, your naming conventions. Loads on customer-facing copy. Often the easiest first skill because the rules already exist on a wiki — you're just porting them.

Archetype 02

The runbook.

A procedure for a recurring task: "how we write ADRs," "how we draft a security review," "what we check before merging a migration." Bundles templates the agent fills in.

Archetype 03

The convention pack.

How your team does something the rest of the world does differently — your testing framework choice, your folder layout, your preferred CI patterns. Stops the agent from inventing a parallel universe.

Three parts, one takeaway.

Part 01 — a skill is a folder of conditional expertise. The agent only loads its body when your request matches the description in the frontmatter.

Part 02 — the agent finds your skill by reading descriptions, matching against your request, and loading the body. Write the description like a search query; write the body like a runbook; test in a loop.

Part 03 — skills carry knowledge; agents do work. Prefer a skill when the host agent can already do the task and just needs to know how. Reach for an agent when you need a new worker with its own tools and lifecycle.

You don't need to write a perfect skill to start. You need to write a small one, watch what the agent does with it, and iterate. Every skill in production started as a 30-line file that fired on the wrong prompts until its author tightened the description.

Skills are how the agent stops being a stranger to your team. You're not training a model — you're handing it the team's notes. — End of series · go write one
Closing · End of series

Pick the smallest piece of knowledge your agent keeps forgetting. Write that skill first.