Claude Skills explained: teach Claude how your firm works

Last updated: 23 Sep 2026.

If you have used Claude for a while, you will know it can write, summarise and reason well out of the box. What it does not know is how your business does things: the way you structure a proposal, the tone your brand insists on, the checklist your team follows before anything goes to a client. Every time you want that, you have to explain it again. Claude Skills solve exactly that problem. A Skill is a reusable, installable package of your know-how that Claude loads when it is relevant, so the right procedure and the right standards turn up automatically rather than living in someone's head or in a prompt they have to paste in every time.

This article explains what Claude Skills are in plain terms, where they work, how a business creates or commissions one, what to check before installing a Skill someone else wrote, and where Skills pay off. It is written for decision-makers rather than developers. Skills launched in Oct 2025 and are now an open standard that other AI vendors support, so this is an established option your operations lead should be weighing up rather than a new feature to watch.

What a Claude Skill actually is

The simplest way to think about a Skill is a folder of instructions and supporting material that teaches Claude how to do a specific job the way you want it done. Inside that folder you might have a written procedure (how to build a proposal, step by step), reference documents (brand guidelines, an approved pricing table, a clause library) and any templates Claude should follow. Technically, the folder contains one required file called SKILL.md, holding a short name, a description of when the Skill applies and the instructions themselves, with reference files or small scripts alongside. Anthropic's docs describe the result as "organized like an onboarding guide you'd create for a new team member". You package it once. From then on, Claude recognises when a task matches the Skill and pulls in those instructions on demand.

That "on demand" part is why a firm can install dozens of Skills without slowing Claude down. At the start of a conversation Claude sees only each Skill's name and description. It reads the full instructions when a request matches, and opens reference files only if the task needs them. The important word is reusable. A clever one-off prompt helps the person who wrote it, on the day they wrote it. A Skill captures that expertise so anyone on the team gets the same result, repeatedly. Skills also sit naturally alongside Claude AI agents for business: an agent decides what to do and takes action, and a Skill is the packaged expertise it draws on to do a particular job to your standard.

How Skills differ from a prompt, a Project or a connector

Anthropic's own help centre draws the lines this way:

• A prompt is instructions for a single conversation: not saved, not shared, not consistent across your team. Skills are installed once and available every time.
• Custom instructions are a standing preference applied to everything. In Anthropic's words, "Custom instructions apply broadly to all your conversations. Skills are task-specific and only load when relevant, making them better for specialized workflows."
• A Project (the kind you build with Claude Projects) is reference material. Anthropic puts it as: "Projects provide static background knowledge that's always loaded when you start chats within them. Skills provide specialized procedures that activate dynamically when needed and work everywhere across Claude." That is the difference between what Claude knows and how Claude does a task, and the two complement each other.
• A connector to your CRM or document store is about what Claude can reach, handled through approaches like the Model Context Protocol (MCP). Anthropic's help centre says MCP connects Claude to external services and data, while Skills supply the procedural knowledge for completing a task. A Skill can tell Claude how to use a connection, but it is not the connection itself.

Skills are now an open standard

This is the change most worth knowing about since we first published this article. Anthropic released the Skills format as an open standard on 18 Dec 2025. The specification lives at agentskills.io, which describes it as "originally developed by Anthropic, released as an open standard". A Skill you write is a folder of plain text files in a documented format, and other vendors' tools can read it.

Several already document that support. OpenAI's guide for Codex and ChatGPT states that "Skills build on the open agent skills standard" and uses the same SKILL.md file. GitHub's documentation says "The Agent Skills specification is an open standard, used by a range of different AI systems", and Copilot reads Skills from a .claude/skills folder as well as its own. Cursor loads Skills from Claude and Codex folders "for compatibility", and Google's Gemini CLI describes its Skills as "Based on the Agent Skills open standard".

Two honest caveats. That list is dominated by developer tools: no CRM, finance system or office suite has adopted the standard, so the Skill your marketing team uses in the Claude app will not run inside those platforms today. And each vendor implements the standard in its own product, so a Skill that leans on a Claude-specific feature may need adjusting elsewhere. Even so, the work of codifying your procedures is portable know-how rather than a bet on one model.

Where Skills work, and where they do not carry over

Anthropic supports Skills across the Claude apps, Claude Code, the Claude Agent SDK and the Claude Developer Platform (the API firms use when they build Claude into their own software).

Anthropic's documentation is clear that custom Skills "do not sync across surfaces". A Skill uploaded to the Claude app must be uploaded separately to the API, and Skills in Claude Code live in files on the developer's machine, separate from both. Keep a single master copy under version control and treat each upload as a deployment. Uploading your own Skills to the Claude app needs a Pro plan or above with code execution enabled, and by default those Skills belong to the person who uploaded them. On Team and Enterprise plans you can share a Skill with colleagues, who receive it as view-only. Organisation-wide distribution is a separate admin feature, covered below.

Concrete business examples

Skills are easiest to grasp through jobs your team already does. None of these require writing software. They require codifying expertise you already have.

A Skill that enforces your proposal format

Most firms have a proposal structure that works: a particular order of sections, a scope statement written a certain way, a consistent pricing table, the right terms and disclaimers. Quality drifts depending on who writes it and how busy they are. A proposal Skill packages the structure, the standard wording and the must-include clauses, so when someone asks Claude to draft a proposal, the review is about the deal rather than the formatting.

A Skill that holds your brand tone

Brand voice is one of the first things to slip when several people write content. A tone-of-voice Skill captures how your organisation writes: the words you use and avoid, formality, how you talk about competitors, whether you use British spelling. Apply it to a blog draft, a sales email or a customer reply and the output sounds like your business rather than a generic assistant.

The same pattern covers monthly reporting in your standard format, compliance checks against a fixed checklist before anything is sent, support replies that follow your escalation rules, and tender responses assembled from an approved answer library. Wherever there is a repeatable job with a right way of doing it, a Skill captures that way so the quality does not depend on who is at the keyboard.

What early adopters have said

Independent case studies on Skills are still thin, so treat the following for what it is: quotes published by Anthropic in its own launch announcement on 16 Oct 2025, from launch partners rather than an independent evaluation. The most concrete is from Rakuten's General Manager for AI, Yusuke Kaji: "Skills streamline our management accounting and finance workflows. Claude processes multiple spreadsheets, catches critical anomalies, and generates reports using our procedures. What once took a day, we can now accomplish in an hour." Box described turning stored files into documents that follow the organisation's standards, and Notion described more predictable results with less prompt wrangling. In each case it is a procedure the organisation already had, applied consistently.

How a business creates or commissions a Skill

You do not need to be technical to start, but you do need clarity about the job. Most of the effort goes into capturing expertise precisely.

1. Pick a job that is repeatable and standardised. Tasks your team does often, that follow a known method, and where inconsistency currently costs time or quality.
2. Write the procedure down. The steps, the standards, the do's and don'ts. If a senior colleague does this job well, capturing how they do it is the heart of the Skill.
3. Gather the supporting material. Templates, approved wording, brand guidelines, example outputs you are happy with.
4. Build and test it. Check it produces the standard you want, triggers when it should and stays out of the way when it should not.
5. Roll it out and maintain it. Decide who owns it and refine it as your standards change. A Skill is a living asset, not a one-off.

Simpler Skills can be built in-house once someone has learned the format. For Skills that touch sensitive material, connect to other systems or need to behave reliably across a whole team, most businesses commission help so the result is robust and properly tested. That is the work SpotDev does: we are an in-house, fully remote engineering team, we are Certified in the Claude Partner Network, and we run our own marketing operations on Skills we wrote ourselves. To scope a Skill or a wider rollout, talk to a Claude-specialist engineer through our Claude implementation packages.

Before you install someone else's Skill

Because a Skill can contain instructions and scripts, installing one is closer to installing software than to saving a document. Anthropic's guidance is blunt: "Use Skills only from trusted sources: those you created yourself or obtained from Anthropic." Its docs warn that "a malicious Skill can direct Claude to invoke tools or execute code in ways that don't match the Skill's stated purpose".

The risk is not theoretical. In Feb 2026 the security firm Snyk scanned 3,984 Skills from two third-party marketplaces, ClawHub and skills.sh, and found critical security flaws in 13.4 per cent of them (534 Skills), including 76 it confirmed as malicious, built to steal credentials or send data elsewhere. Those marketplaces are not run by Anthropic, but anything downloaded from an open directory needs the scrutiny you would give an unknown browser extension. Before installing a Skill you did not write, have someone check four things:

1. Can you read every file? Open the SKILL.md and every bundled script and reference file. If you cannot read it all, do not install it.
2. Does it reach outside itself? Look for anything that fetches web pages, calls external addresses or reads files beyond its own folder. Anthropic flags Skills that fetch external content as a particular risk, because that content can carry instructions of its own.
3. Do the instructions match the stated purpose? A brand-tone Skill has no business asking Claude to run commands, hide its actions or ignore other rules.
4. What can Claude reach while it runs? Be strictest where Claude is connected to customer data, finance systems or anything that can send email.

We cover ownership, review, versioning and retirement of Skills across a firm in how to govern Claude Skills, and the wider question of accountability for AI decisions in our AI governance framework for UK businesses.

Admin controls on Team and Enterprise plans

On Team and Enterprise plans, organisation owners can provision approved Skills to everyone in the organisation from one place, switch user-created Skills off entirely, and set a publishing policy that requires an owner's review before a Skill goes live. Enterprise organisations can also turn on automated scanning of Skills uploaded in the Claude app: a Skill that fails, or has not yet been scanned, is blocked. Two limits: scanning does not cover Skills uploaded through the API, and the audit log records that a Skill was shared and with whom, not what it contained. Central provisioning is what turns a good Skill written by one person into the standard for the whole business.

Where Skills pay off, and where they do not

Skills are worth the investment when the same expertise needs to be applied again and again, consistently, by more than one person. The return shows up as less rework, faster turnaround on standard documents, fewer errors slipping through, and senior people freed from re-explaining the basics. They are less useful for one-off tasks, or for work that is different every time. The honest test: is there a method here that you would want every member of the team to follow? If yes, it is a strong candidate. If the value is in fresh thinking each time, a Skill adds little.

Frequently asked questions

Do I need to be a developer to create a Claude Skill?

No. A Skill is a folder of plain text instructions and reference files, and the hard part is capturing the procedure and standards clearly, which is a business task rather than a coding one. Simpler Skills can be assembled in-house. For Skills that need to be reliable across a whole team, connect to other systems or handle sensitive material, most businesses commission an engineering team to build and test them properly.

How is a Claude Skill different from a Claude Project?

A Claude Project holds shared knowledge, the reference material Claude reads to answer questions, and it loads in full whenever you work inside that Project. A Skill holds method, the steps and standards for doing a specific task your way, and it loads only when a task calls for it. They work well together: the Project holds the knowledge and the Skill enforces how the work is done.

Do Claude Skills work with other AI tools?

Increasingly, yes. Anthropic released the Skills format as an open standard in Dec 2025, and OpenAI's Codex, GitHub Copilot, Cursor and Google's Gemini CLI each document support for the same SKILL.md format. Support is strongest in developer tools, and business platforms such as CRMs have not adopted it.

Is it safe to install a Skill someone else wrote?

Only if you have read it. Anthropic's guidance is to use Skills only from trusted sources, meaning ones you wrote or obtained from Anthropic, because a Skill can contain scripts and instructions that do more than its description claims. Before installing, read every file, check whether it reaches outside its own folder, confirm the instructions match the stated purpose, and limit what Claude can access while it runs.

Work with a Claude specialist

SpotDev designs, builds and deploys custom Claude agents, Skills and enterprise Claude rollouts for UK businesses, with fixed packages from £8,000 to £45,000 and a first rollout live in two to three weeks. Explore our Claude implementation packages or request a quote for a Skills rollout scoped to your procedures.