What are Claude skills, and which ones are worth having?

Every Monday, an operations analyst produces the same variance report. Round to the nearest thousand. Flag anything more than 5% off budget in amber, more than 10% in red. Put commentary under each region, never in a separate section. Use the finance team's names for cost centres, not the ERP codes.

She could paste those rules into every chat. She could put them in a project, where they'd sit in context even when she's asking about something else. Or she can write them down once as a skill, and Claude will load them when, and only when, she asks for the variance report.

Anthropic defines skills as "folders of instructions, scripts, and resources" that Claude loads only when relevant (Anthropic Help Centre, What are skills). The loading is the point. A skill stays out of the way until a task matches it, which keeps Claude's context small. Twenty installed skills cost almost nothing until one is needed.

There are four kinds:

Skills launched in October 2025 for paid plans. In December 2025 Anthropic added organisation-wide management and a partner directory, and published Agent Skills as an open standard (Anthropic, Introducing Agent Skills). The help centre now lists them on Free as well.

Installing an existing skill

1. Turn on code execution (Settings > Capabilities). Skills need it, because many of them run scripts.
2. Open Customize > Skills, select "+", then "Browse skills".
3. Switch on the ones you want. On Team and Enterprise, admins can provision skills for the whole organisation.

Writing your own

A custom skill is Markdown instructions, with scripts as an optional extra. The instructions need two things: a description precise enough that Claude knows when to use the skill, and steps precise enough that two runs produce the same thing.

The core of the variance-report skill (illustrative):

Use this skill when asked for the weekly variance report or a budget-vs-actual
summary by region.

1. Round all figures to the nearest £1,000.
2. Variance = actual minus budget. Variance % = variance divided by budget.
3. Flag |variance %| > 5% as amber, > 10% as red.
4. Use the cost-centre names in names.csv, never ERP codes.
5. Under each region, write two sentences of commentary. Name the line
   items that explain most of the variance, with their figures.
6. End with a list of any rows you excluded and why.

Note step 2. "Variance" can be read as budget minus actual or the other way round, and a model will pick one without saying so. A skill is where you settle that question for good.

Developers can use the same skills in Claude Code and through the API (in beta, with the code execution tool and the /v1/skills endpoint).

A skill tells Claude to follow your method. To see whether it did on a given answer, connect Count: it shows the reasoning chain behind each claim, so a skipped step is visible.

Lists of "best Claude skills" tend to feature clever demos. For analysis, finance and operations work, we'd rank them by a duller test: how often would you otherwise retype the same instructions, and how bad is it when Claude improvises?

Start with Anthropic's document skills. The Excel, Word, PowerPoint and PDF skills are what let Claude produce real files rather than text that looks like a table. If you hand work to people who live in Office, these do most of the lifting.

Then write the skills only your team can write. Candidates that pay back quickly (illustrative):

Use partner skills where you already use the product. If your team lives in Notion or Jira, the vendor's skill encodes how its tool expects to be used.

The counter-argument is fair: for one-off tasks, a good prompt is quicker than a skill. Skills earn their keep on the tenth run, not the first.

Anthropic's help centre puts the distinction in one line each: projects hold always-loaded, static knowledge; skills are loaded on demand; MCP connects Claude to services, while skills teach it how to use them.

In practice the three combine. A finance team might keep its board-pack project for this quarter's documents, connect its warehouse through a connector, and write a skill that says how to turn warehouse numbers into the board's format. The project knows what's true this quarter; the connector knows today's numbers; the skill knows how the team works.

If you're choosing just one: a rule that applies to every chat in one piece of work belongs in a project. A procedure you'd use across many pieces of work belongs in a skill.

Skills can execute code, and Anthropic's advice is direct: "stick to trusted sources to keep your data safe" (Anthropic, Introducing Agent Skills). Enterprise admins can switch on scanning of skills and plugins for malicious content.

The risk is the same one that has already bitten MCP users. In April 2025, Invariant Labs showed that hidden instructions in a tool's description could lead an agent to exfiltrate SSH keys and config files, and that a description could change after a user had approved it. A skill's instructions are read by Claude as guidance, which makes them a natural place to hide the same kind of instruction.

Code execution brings its own warning. Anthropic says the sandbox "gives Claude internet access to create and analyze files, which may put your data at risk", and that prompt injection through files or websites could lead to data leaving the conversation (Anthropic Help Centre, Create and edit files). Before installing a skill from outside your organisation, read its instructions and scripts. It takes five minutes, and it's the only review it will get.

Nobody has published accuracy figures for Claude skills, and Anthropic documents no numeric limits for them. What we can say from how they work: a skill makes Claude repeat a procedure faithfully. If the procedure has a mistake in it, Claude repeats the mistake faithfully too, in every report, for everyone the skill is provisioned to.

There's also a quieter failure. Skills load only when Claude judges them relevant. Ask for "the Monday numbers" instead of "the weekly variance report" and the skill may not load at all, leaving Claude to improvise the format it was meant to settle.

A two-minute check for any skill

1. Ask which skills it used: "Which skills did you use for that answer?" If the one you expected isn't named, reword your request or sharpen the skill's description.
2. Run a known case. Keep one input where you already know the right output (last month's report, checked by hand). Run it after every edit to the skill.
3. Check each rule. Go down the skill's numbered steps and tick them off against the output. Rounding and sign conventions are the usual misses.
4. Read the excluded rows. If your skill asks Claude to list exclusions, read the list. That's where a bad filter shows up.

Steps 2 to 4 check the procedure. They don't check the claims inside a particular answer. Count does that part: connect it to Claude, ask it to audit an answer, and each claim comes back labelled as a citation, an assumption, model knowledge or a hallucination, with the reasoning chain behind it.

What are Claude skills?

What are the best Claude skills?

What are Claude Agent Skills?

What's the difference between Claude skills and MCP?

Are Claude skills free?