Claude Code Skills for Agencies: The 2026 Guide
How agencies write Claude Code skills that cut hours off repeat work: the SKILL.md format, install, team sharing, and versioning.
The weekly client report eats half your Friday. Same format every week, rebuilt from scratch every time. That is a skill waiting to be written.
Claude Code skills are folders with a SKILL.md file that teach Claude a task, from code review to client reporting. Agencies use them to cut hours off repeat work and package expertise the whole team reuses. Skills live in ~/.claude/skills/ and travel with git.
Most writing on Claude Code skills is a copy of the docs, or a list of templates nobody ran. This is the plain version. It comes from a team that runs skills on client work every week.
TL;DR
- A Claude Code skill is a folder with a SKILL.md file that teaches Claude one job. It loads by itself when the task matches its description.
- Write the name and description in YAML frontmatter, the steps in plain English, and one worked example at the top.
- Start with the bundled skills, then write one for the client task you repeat weekly. Keep every skill in git so the team runs one version.
- A skill stops at instructions. Live data comes from a connection. Memory fades between sessions. Judgment stays human.
What are Claude Code skills?
A Claude Code skill is a folder with a SKILL.md file. YAML frontmatter gives the skill a name and a description. The body holds plain instructions for one job. Claude reads the description, and when your request matches, it follows the steps.
Agent Skills is an open standard, built by Anthropic and adopted across many AI tools. Claude Code skills follow that same standard, per the claude code skills docs. Any compatible tool reads it.
Claude loads the name and description at startup. The full file waits for a match. The Agent Skills docs call that progressive disclosure.
Skills load from a few places. The install section below maps them. Personal and project folders are the two that matter for agency work.
A skill fires on its own when the task matches. That automatic trigger is the whole point.
How do you write a Claude Code skill for agency work?
Write a SKILL.md file with two parts: YAML frontmatter that names the skill and describes the job, then plain instructions for the steps. Start with one task you do weekly. Keep the body short and the steps plain.
Name the folder for the job. The folder name becomes the command you type, and the description decides when Claude loads the skill automatically, per the claude code skills docs. Anthropic’s free Claude Code 101 course covers the same ground in lesson 10, per Claude Academy.
The lesson opens with the exact agency pain: you re-explain your standards in every review. A skill stores them once.
Here is a working example. It builds a weekly client report from project notes.
---
name: client-status-report
description: Builds the weekly client status report from project notes. Use when the user asks for a status report, a weekly update, or a client recap.
---
Build the weekly status report for one client.
1. Read the project notes for this client.
2. List what shipped, what is in review, and what is blocked.
3. Write it in plain words the client already understands.
4. Save it as markdown named after the client and the week.
Example: loose notes become one page the client can skim.That file is the whole skill. Claude reads the description, matches it to a request, and follows the steps. A worked example at the top beats extra rules.
Supporting material goes in folders beside the SKILL.md, like scripts and references. Tools a skill may call get pre-approved in an allowed-tools frontmatter field, the claude code skills docs explain. Declare only what it uses.
Our first review skill shipped with rules and no example. It failed on real drafts, every run. The fix: a worked example on top, and rules trimmed to what the example could not show.
I run five skills in this site’s pipeline:
- researcher
- planner
- copywriter
- reviewer
- publisher
Each carries a SKILL.md. This post is their output.
The prompts our team actually runs are packaged as a skills bundle at automatereal.com/skills. All skills sell for $99.
AutomateReal skills
The example matters more than the rules. Claude does better when it sees one good result.
Which Claude Code skills should an agency install first?
Start with the bundled skills that ship with Claude Code. They cover the work agencies repeat most. Then pull one community pack, and write your own skill for the client job you do every week.
The bundled set ships with Claude Code. It is prompt-based, meaning Claude gets detailed instructions and does the work with its own tools, per the claude code skills docs. These six map cleanly to agency work.
| Skill | What it does | Install it when |
|---|---|---|
| /doctor | Checks the Claude Code setup | Day one |
| /code-review | Reviews changes against your standards | You ship work to clients |
| /batch | Runs one prompt across many files | You have a repeat workflow |
| /debug | Walks through a failing build or test | Something breaks |
| /loop | Repeats a task until a success check passes | A job runs long |
| /claude-api | Turns an API spec into working code | You build integrations |
Community packs like Superpowers and Claude-Mem ship ready-made skills. GSD covers task management. Our best Claude Code skills rundown ranks the packs we tested.
Treat them as samples. Audit each file before it joins the team folder.
Match beats popularity. The skill for your agency is the one tied to a job you did by hand last week.
How do you install and share skills across a team?
Copy the skill folder into ~/.claude/skills/ for your own setup, or into .claude/skills/ inside the project for the team. Anyone who clones the repo gets the project skills automatically. That is the whole install.
Personal skills apply to every project. Project skills live in .claude/skills/ at the repo root, and committing them shares them with the team, the claude code skills docs explain. One pull sets up a new hire, which is the whole onboarding.
Every skill that lands in my folder passes one test. I read the file, and I can still explain the skill in one sentence. If I cannot, it does not ship.
Install is one folder copy. Sharing is one git repo.
Version skills like you version code
Every skill change goes through git. One commit per edit, one review before merge. One re-test on real work after the pull.
A skill is instructions, and instructions drift. Someone tightens a description, and the skill starts firing on half your sessions. Someone edits a step, and the output changes shape.
The team file lives in one repo. Keep it private when it holds client playbooks.
Our review skill went through this loop. One commit and a second review, then a re-run on a real draft before it earned its place.
Team skills drift. The version history is how you find out why.
How is a skill different from a prompt or a slash command?
A prompt is a one-off instruction you type, and a slash command is a shortcut you type by name. A skill is a folder that teaches Claude a whole job and loads on its own when the task matches. They share a family, but only a skill reloads itself.
Custom commands have merged into skills. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way, the claude code skills docs confirm. The docs recommend a skill for new work, because skills support supporting files.
Skills sit next to two other pieces: subagents and MCP servers. The table tells them apart.
| Piece | What it is | When to use it |
|---|---|---|
| Prompt | A one-time instruction you type | A task you do once |
| Slash command | A shortcut with fixed behavior | Built-in checks like /doctor |
| Skill | A folder of instructions that loads by description | A job you repeat weekly |
| Subagent | A separate Claude run with its own context | Parallel work that should not touch your history |
| MCP server | A connection to live tools and data | A job that needs current information |
Skills and subagents work together. Add context: fork to a skill’s frontmatter. The skill runs inside a subagent, isolated from your conversation history, per the claude code skills docs. The subagent gets the skill content as its prompt.
That separation matters for agency work. The skill holds the standard. The subagent holds the isolation.
Prompts and commands give instructions on call. A skill reloads by itself, and it travels with the team.
Where skills stop being enough
A skill carries instructions. Live data comes from a connection, and memory fades between sessions. Judgment stays human.
The last one is the agency rule. A skill drafts. A person decides tone and scope, and the client output gets a final read.
My own outreach lands ~200 form submissions a day, and automated follow-up handles the replies. The skill drafts the message. Deciding which reply deserves a real call still lands on me.
When a job outgrows a skill, the right part plugs in around it.
- MCP server: the live data connection.
- Subagent: an isolated run that stays out of your history.
- Plugin: wraps the skill with agents, hooks, and MCP servers, as the claude code skills docs note.
Plan the human step. Client-facing output gets a final read by a person, every time.
FAQ
What is the SKILL.md format for Claude Code skills?
SKILL.md is a markdown file with YAML frontmatter at the top. The frontmatter holds the name and description. The body holds the plain instructions, with optional folders beside it.
Where are Claude Code skills stored on disk?
Personal skills go in ~/.claude/skills/. Project skills go in .claude/skills/ at the repo root. Both are plain folders you copy or delete.
Do skills work with Claude Code subagents?
Yes. Add context: fork to the frontmatter, and the skill runs inside a subagent, isolated from your conversation history. The skill content becomes the prompt that drives the subagent.
How do I write a skill that calls a tool or a script?
List the tools in an allowed-tools frontmatter field, then describe the call in the body. Skills can bundle scripts and run them when matched. Test against the real tool.
Your first skill this week
Pick one client task you did by hand last week, and write the skill for it. Run it once on real work. Keep it only if it beats your hand version.
The format is new enough that nobody has ten years of it. Trust the skill you can read. If you want help finding the first workflow worth a skill, a discovery call maps it in about 30 minutes.
AutomateReal services