Claude skills are packaged instructions that teach Claude how to do a specific job your way. If you have ever explained the same task to an AI three times, with the same caveats and the same output format, a skill is how you stop explaining. You write the process down once, Claude loads it when the task comes up, and the result comes out consistent.
We run our own business on a set of skills: one reconstructs monthly billable hours, one audits Amazon ad accounts, one runs a cross-channel marketing review, and one makes drafts sound like the person who will send them. This guide explains what skills are, how they work under the hood, and how to decide which of your processes should become one.
What a skill is
A skill is a folder. Inside the folder is a file called SKILL.md, which holds a short header and the instructions. The folder can also hold anything else the job needs: scripts that do calculations, reference documents, templates, examples of good output.
The header has two required fields:
---
name: amazon-ads-audit
description: Audits an Amazon Advertising account and produces a prioritized list of fixes plus an Excel workbook. Use when asked to audit, review or find wasted spend in an Amazon ads account.
---
The name is a short identifier. The description is the most important sentence in the whole skill, because it is how Claude decides when to use it. It should say both what the skill does and when it applies.
Below the header is the body: the actual procedure, written in plain language. Which data to pull, in what order, what to check, what thresholds matter, how to format the output, and what never to do.
Anthropic originally developed the format and released it as an open standard called Agent Skills, so the same folder structure now works in a number of AI tools beyond Claude.
How Claude uses skills without drowning in them
The clever part of skills is what Anthropic calls progressive disclosure. It works in three stages:
- Discovery. At the start of a session Claude loads only the name and description of each available skill. That is a few lines per skill.
- Activation. When your request matches a description, Claude reads the full SKILL.md instructions.
- Execution. Claude follows the instructions, opening reference files or running bundled scripts only when a step calls for them.
This is why you can install dozens of skills without slowing anything down. Claude is not reading your entire operations manual on every request. It reads the table of contents and opens the chapter it needs.
Skills versus prompts, projects and connectors
People mix these up, so here is the distinction we use:
| Tool | What it does | Analogy |
|---|---|---|
| A prompt | Tells Claude what you want right now | A verbal request |
| Project instructions | Standing context for one workspace | A briefing for one client file |
| A connector (MCP server) | Gives Claude access to a system | Keys to the building |
| A skill | Tells Claude how to do a specific job | The written procedure |
Connectors and skills work together. A connector to Amazon Ads lets Claude pull campaign data. A skill tells it which reports to pull, which thresholds signal waste, and how to lay out the workbook. Access without a procedure produces inconsistent answers. A procedure without access means you paste data in by hand. If you are new to connectors, start with what MCP servers are.
Four skills we use, and what each one taught us
Monthly hours reconstruction
Every month we need an hours log for client billing. Nobody keeps a perfect timesheet. The skill pulls three evidence sources, sent email, Google Drive file activity and the calendar, and reconciles them into one line-item spreadsheet with date, task, type, project and hours. The methodology lives in the skill, so it is applied the same way every month instead of being reinvented.
Lesson: skills are ideal for work where the method matters as much as the result. A consistent method is what makes the output defensible.
Amazon ad account audit
This skill connects to the Amazon Ads account, checks Sponsored Products, Sponsored Brands and Sponsored Display, and returns a prioritized list of fixes with specific campaign names and spend figures, plus a multi-tab Excel workbook. It knows what to look for: search terms spending without converting, campaigns capped by budget, missing negatives.
Lesson: a skill is where you encode expertise. The checks in that audit took years to learn. Now they run in minutes.
Cross-channel "360" marketing review
For one brand, this skill looks across Meta ads, Google Ads, Shopify, GA4 and Search Console and returns one combined list of action items covering wasted spend, conversion issues and technical SEO problems.
Lesson: skills shine when a task spans several systems that no single vendor dashboard covers.
Writing in the owner's voice
We built a writing-style skill from 120 days of the owner's sent mail. It captures how he opens, how long his emails run, which phrases he uses and which he never would. Drafts now need light edits instead of rewrites.
Lesson: some of the most useful skills are not technical at all. They capture taste.
How to tell which processes should become skills
A process is a good skill candidate when most of these are true:
- You do it at least monthly.
- You have explained it to a person or an AI more than twice.
- There is a right way to do it, and deviations cause problems.
- The output has a standard format: a spreadsheet, a memo, an email, a report.
- It requires judgment along the way, not just data movement.
That last point matters. If the task is purely "move this data from A to B at 6:00 every morning," it belongs in a scheduled workflow, not a skill. We explain that split in n8n vs Claude for workflow automation.
Building your first skill
You do not need to be technical. Here is the process we follow:
- Do the task once with Claude, out loud. Walk through it in a normal conversation, correcting Claude as you go.
- Ask Claude to write the procedure. At the end, ask it to turn what you just did, including your corrections, into a SKILL.md. Claude is good at this, and Anthropic provides a skill for creating skills.
- Edit the description hard. Make it specific about when to use the skill, including the phrases people actually say: "run hours," "audit the ads," "do a 360."
- Add the rules you care about. Output format, what to leave out, where to save results, what requires approval. Be explicit. "Never send; draft only" belongs in the skill, not in your memory.
- Test on a real case. Run it on last month's data and compare to what you would have produced.
- Revise after each use. Every correction you make is a line that belongs in the skill. Good skills get better every month.
Where skills work
Skills are available in the Claude apps, in Claude Code and through the Claude API. In the Claude apps, custom skills are available on Pro, Max, Team and Enterprise plans with code execution turned on, and you upload a skill as a zip file in settings. In Claude Code, skills live in a folder on disk, either personal or inside a project. Anthropic also ships built-in skills for working with Word, Excel, PowerPoint and PDF files.
One caveat: custom skills do not sync automatically between these surfaces. If you use the same skill in the app and in Claude Code, plan to keep both copies current, or keep the master in one shared location.
Security and governance
A skill is instructions plus, sometimes, code. Treat it like software you install. Anthropic's guidance is to use skills only from trusted sources, meaning ones you wrote or ones from Anthropic, and to audit every file in a skill you did not write. A malicious skill can direct Claude to misuse its tools.
For a company, that means a few simple rules. Keep skills in version control or a shared, permissioned folder. Have one owner per skill. Review changes the way you would review a change to a standard operating procedure. And keep human approval steps written into any skill that sends, publishes or spends.
What this looks like in a business
The end state is not "AI does everything." It is that your best person's way of doing a job is written down, runs the same way for everyone, and improves every time someone corrects it. A new hire can run the monthly close review on day one because the method lives in the skill, not in someone's head.
That is the core of our AI workflow automation work: finding the processes that should be skills, writing them with the people who do the work, and connecting them to the systems they need.
Frequently asked questions
Do I need to know how to code to build a Claude skill?
No. The core of a skill is a plain-language document. Scripts are optional and only needed for steps like precise calculations or file conversions. Claude can draft the skill for you after you walk it through the task once.
How is a skill different from a saved prompt?
A saved prompt is text you paste in. A skill is loaded automatically when Claude recognizes the task, can bundle reference files and scripts, and only loads its full contents when needed. Skills also scale better, since dozens can be installed without filling up every conversation.
Can my whole team share the same skills?
Yes, with some care. In Claude Code, project skills live in the project folder and are shared with anyone who uses that project. Skills uploaded through the API are shared across the workspace. Skills uploaded in the Claude app settings are per user, so teams usually keep a master copy in a shared location.
How long does it take to build a useful skill?
A first version often takes an hour or two, most of it spent doing the task once with Claude and tightening the description. The real value comes over the next several uses, as you add each correction to the skill.
Will a skill always produce exactly the same output?
It will follow the same method, but Claude still uses judgment, so wording and emphasis can vary. For steps that must be exact, such as calculations or file formats, put the logic in a bundled script so that part is deterministic.