Most AI disappointment in a small business has the same cause. The model is capable, but it has never been told how this particular firm does things. It does not know that quotes always carry a fourteen day validity, that the Leeds site invoices separately, or that you never chase a retainer client on the same day as a project client. So it produces something generic, someone rewrites it, and the conclusion is that AI does not really work here.
Agent Skills are the mechanism Anthropic built for exactly that gap.
What are Claude AI Skills, and how do they differ from prompts?
The difference matters more than it first sounds. A long prompt is something you paste, use once, and lose. A Skill sits on disk, is version controlled if you want it to be, and activates itself.
Anthropic's Agent Skills documentation compares the structure to an onboarding guide you would create for a new team member, which is the right mental model. You are not writing a program. You are writing the note you wish you had given the last person who joined.
Why do most small businesses get worse AI output than they should?
That gap between trying and embedding is where Skills earn their keep. A firm using a chat window with no persistent context is competing on prompt writing talent. A firm with five well written Skills is competing on process quality, which is a much more defensible thing to be good at.
What actually goes inside a SKILL.md file?
A first Skill for a UK trades business might be named quoting-domestic-jobs, with a description saying it produces customer facing quotes for domestic installation work and should be used whenever the user asks for a quote, estimate or price breakdown for a residential job. The body then carries the actual rules: the standard validity period, how VAT is presented, what the call out charge covers, which materials are quoted at cost plus and which are fixed, and the sentence you always include about access requirements.
None of that is clever. All of it is the stuff a new estimator gets wrong for three months.
How does Claude decide which Skill to use?
The consequence for a small business is that installing more Skills is close to free. Anthropic notes there is no practical limit on bundled content, because files do not consume context until accessed, and that scripts run through bash so only their output enters the context window, never the code itself.
What should a UK small business build a Skill for first?
A reasonable first three for a UK services firm:
- Enquiry triage and first response. Your tone, your qualifying questions, your standard next step, and the rule about what never goes in a first reply.
- Quote or proposal generation. Structure, pricing presentation, VAT handling, validity, exclusions. This overlaps heavily with the ground covered in writing quotes and proposals with AI.
- Recurring paperwork. Whatever your monthly cycle actually involves, written as a numbered workflow.
If you have never written your processes down at all, that comes first. There is a genuine dependency here, and it is worth reading up on documenting business processes before automating them rather than trying to invent the process while writing the Skill.
How do you write a Skill that Claude actually triggers?
The most useful concept in the Skill authoring best practices guide is degrees of freedom. Anthropic frames it as a robot on a path. Where there is one safe route and cliffs on both sides, give exact instructions and no room to improvise. Where the field is open and many routes work, give general direction and trust the model.
Applied to a UK business: the sequence for submitting a VAT return is a narrow bridge, so write it as an exact numbered procedure with no optional steps. Drafting a reply to an awkward customer email is an open field, so give principles and tone rather than a script.
Two further rules worth obeying literally. Keep file references one level deep from SKILL.md, because Anthropic warns that Claude may only partially read files reached through nested references. And for any reference file longer than 100 lines, put a table of contents at the top so the full scope is visible even on a partial read.
Where do Skills live, and do they follow you between Claude apps?
The sharing model differs too, which catches teams out. On claude.ai, Skills are individual to each user, each team member uploads their own, and there is no centralised admin management or organisation wide distribution. On the API they are workspace wide and all workspace members can access them.
There is also a runtime difference that shapes what a Skill can do. Skills running through the Claude API sit in a sandboxed container with no network access and no runtime package installation. Skills running in Claude Code have the same network access as any other program on the machine. A Skill that needs to call your booking system will work in one place and silently fail in the other, so decide the surface before you design the Skill.
Practical answer: keep the master copy in one versioned location and treat every surface as somewhere you publish to. It is the same discipline as keeping one source document rather than four slightly different copies on four laptops.
How do you test a Skill before trusting it with real work?
That ordering matters and most people get it backwards. Writing the Skill first produces a document full of guidance for problems that never actually occurred, which costs context and buries the rules that do matter.
Anthropic also describes a two instance loop worth copying: work with one Claude session to design and refine the Skill, test it with a separate fresh session doing real tasks, then bring specific observations back to the first. Watch what the testing session ignores. If it never opens a bundled file, that file is either unnecessary or badly signposted from SKILL.md.
What are the security and governance trade offs?
For a UK firm holding client data, the review process is straightforward but non negotiable. Read every bundled file, not just SKILL.md. Look at the scripts. Check for network calls that have nothing to do with the stated job. Anthropic's own framing is to treat it like installing software, which is exactly the right level of caution for a downloaded Skill from an unknown author.
Worth noting for anyone in a regulated line of work: check Anthropic's current data retention terms for Agent Skills against your own obligations before a Skill carries anything client identifying. If you advise clients on compliance, as covered in the guidance on AI for accountants and solicitors, that check belongs in the same review as the file read.
What does a realistic first month look like?
Resist building the library up front. A Skill written from memory of how a process works is usually wrong in the details, and the details are the entire point. A Skill written the day after you did the task by hand tends to be right, because the friction is still fresh.
Expect the biggest surprise to be organisational rather than technical. Writing down how the business quotes, or invoices, or handles a complaint frequently surfaces the fact that two people do it differently. Skills do not fix that. They just make it visible, which is what any serious look at where AI genuinely reduces admin hours tends to conclude anyway.
The underlying shift is small and specific. Instead of a general purpose model that knows a lot about the world and nothing about your firm, you get one that has read your quoting standard before it writes your quote. That is not a leap in capability. It is a leap in relevance, and relevance is the thing that was missing.
Sources referenced in this article: Anthropic's Agent Skills overview, Anthropic's Skill authoring best practices, and Anthropic's engineering post Equipping agents for the real world with Agent Skills.
Claude AI Skills — FAQ
What is a Claude Skill in plain English?
A Skill is a folder on disk containing a file called SKILL.md, plus any reference documents or scripts you want to go with it. The SKILL.md file starts with a short block of YAML naming the Skill and describing when it should be used, then continues with ordinary markdown instructions. Anthropic's documentation describes the folder as being organised like an onboarding guide you would write for a new team member. That comparison is the useful one. If you would hand a new hire a two page note explaining how your firm quotes a job, that note is a Skill. Claude reads it only when the work in front of it matches the description, so the guidance arrives at the moment it is relevant rather than being pasted into every conversation.
How is a Skill different from just writing a long prompt?
A prompt is conversation level and one off. You type it, Claude uses it, and the next conversation starts blank. Anthropic's documentation frames Skills as the alternative to repeating the same guidance across conversations: they are filesystem based and load on demand. The practical difference is context economy. Anthropic's own table puts the standing cost of an installed Skill at roughly 100 tokens, because only the name and description sit in the system prompt until something triggers it. You can therefore install a dozen Skills covering quoting, invoicing, client onboarding and complaint handling without any of them crowding out the conversation. A long prompt gives you one set of rules for one session. A Skill library gives you a set of rules per situation, selected automatically.
Do I need to be technical to write one?
No, though you do need to be able to create a folder and a text file. The minimum viable Skill is a SKILL.md file with two frontmatter fields and some prose. No code is required at all, and Anthropic's best practices guide includes worked examples of Skills that contain nothing but markdown instructions, such as a research synthesis workflow and a style guide compliance loop. The genuinely hard part is not syntax, it is knowing your own process well enough to write it down clearly. Most owners discover mid way through that two people in the business do the same task differently and nobody had noticed. That is a useful finding in its own right, independent of any AI.
Will a Skill I build in Claude Code work in the Claude app?
Not automatically. Anthropic's documentation is explicit that custom Skills do not sync across surfaces. A Skill placed in a Claude Code directory is separate from one uploaded to claude.ai, which is separate again from one uploaded through the API. If you want the same Skill in two places you upload it in both places. There is a further wrinkle for teams. On claude.ai, custom Skills are individual to each user and each team member must upload separately, with no centralised admin management or organisation wide distribution. On the Claude API they are workspace wide. Keep the master copy of every Skill somewhere versioned, such as a shared drive or a git repository, and treat each surface as a distribution target.
Are there risks in downloading Skills other people have written?
Yes, and Anthropic states them directly. Its documentation advises using Skills only from trusted sources, because a Skill gives Claude new capabilities through instructions and code, which means a malicious one can direct Claude to invoke tools or execute code in ways that do not match its stated purpose. The listed risks include data exfiltration, unauthorised system access and tool misuse. Skills that fetch data from external URLs are called out as particularly risky, since fetched content may itself contain instructions. The guidance is to treat installing a Skill like installing software: read every bundled file, including scripts and assets, and look for network calls or file access that has nothing to do with the stated job. For a UK business handling client data, that review is not optional.
How long does it take to see a return on building one?
Building a first Skill from a process you already understand is usually an afternoon, including testing. The return depends entirely on repetition volume. A Skill covering a task done twice a year is not worth the maintenance. A Skill covering the way you write quotes, respond to enquiries or prepare month end paperwork earns back the afternoon quickly because every future run starts from your standard rather than a generic one. A more honest framing than hours saved is variance reduced. The value shows up as fewer outputs that need rewriting because they missed a step, used the wrong terminology or ignored a rule everyone in the office knows but nobody had written down.
