Building Claude Skills and sharing them across your organisation

Your best uses of Claude currently live in a handful of people's heads: one colleague has found the right sequence of instructions to write a sharp scoping note, another knows exactly how to triage support tickets — but that know-how never spreads, and everyone reinvents their own recipe. A Claude Skill solves exactly this problem: it is repeatable business know-how that you "teach" Claude once — instructions, standards, examples, sometimes scripts — and then share across the whole organisation, so quality is consistent and nobody has to rediscover the right prompt. This article explains what a Skill is (an Anthropic product capability), how to design a good one, how to govern and roll out a skills library at scale — and, in all honesty, when a Skill is not justified.

One point of framing: METASENSE is a Creative Tech agency based in Vélizy-Villacoublay, and our standing on this subject is concrete — we use Claude and its Skills ourselves, every day, to produce work. This guide passes on a method you can transpose to your own business functions, with deliberately generic examples: knowledge of your business stays yours, we bring the method and command of the tool.

Worth remembering — A Skill turns a good "private" prompt into a shared asset of the organisation: know-how written once, applied automatically by Claude, and available to everyone. The gain is not magic — it comes from three things: choosing the right case (recurring), writing it well (clear, with examples), and governing it (versioned, distributed, maintained).

Turn your AI know-how into Skills →

This article goes deeper into Stage 3 of our overall method. For the full picture (audit, workflows, Skills, orchestration, cost), see the umbrella guide "How to integrate Claude into your company's processes".

What exactly is a Claude Skill?

A Claude Skill is an organised folder of instructions, standards, examples and sometimes scripts, packaged to give Claude a specialised capability that it applies automatically when the task calls for it. It is an official Anthropic product capability ("Agent Skills"), published as an open standard in October 2025. Where a prompt is a one-off instruction valid for a single conversation, a Skill is reusable know-how that loads on demand, without you having to repeat the same instructions every time.

The most accurate analogy comes from Anthropic's own documentation: a Skill is organised "like an onboarding guide you would write for a new joiner". You write the right way of doing things once; Claude then follows it reliably.

A few canonical definitions, to fix the vocabulary:

• A Skill (Claude) is repeatable know-how, packaged in a folder, that you teach Claude and that it draws on automatically when the context calls for it.
• A prompt is a one-off instruction, valid for a single conversation, neither captured nor shared.
• SKILL.md is the file at the heart of a Skill: a header (the skill's name and description) followed by instructions in plain language.
• A skills library is an organisation's full set of Skills, organised, versioned and governed — its estate of AI know-how.

The difference that matters: a brilliant prompt remains a personal and volatile asset; a Skill is an organisational and durable one. That is the whole point of capturing know-how.

Skill, prompt, workflow: what is the difference?

A prompt is an isolated request, a workflow is a tooled sequence for a business task, and a Skill is the reusable know-how that makes that workflow consistent and shareable. The three do not compete: they nest. You often discover a good sequence through prompts, structure it into a workflow, then fix it as a Skill so the whole team applies it the same way.

In short: the Skill is the captured form of know-how. For the layer below — designing the business workflow itself — see the companion article "Building an AI workflow for each business function with Claude".

How does a Skill work technically (without being a developer)?

A Skill rests on a SKILL.md file and a mechanism called "progressive disclosure": Claude loads only what it needs, when it needs it — which keeps Skills light even in large numbers. You do not need to be a developer to grasp the logic, and that is reassuring for anyone steering adoption at scale.

According to Anthropic's documentation, loading happens on three levels:

1. Level 1 — the identity card (always loaded). The SKILL.md header (the name and the description) costs around 100 tokens per skill. Claude simply "knows" a skill exists and when to use it. That is what allows you to install many skills without weighing down the context.
2. Level 2 — the instructions (loaded on trigger). When your request matches a skill's description, Claude reads the body of the SKILL.md (under 5,000 tokens) — the method, the standards, the good practices.
3. Level 3 — resources and scripts (loaded on demand). Reference files, detailed examples, scripts: Claude accesses them only if the task requires it. A skill can therefore carry heavy documentation at no context cost until it is used.

Structurally, a Skill is simply a folder. At minimum, a SKILL.md with a header (the name — lower case, hyphens allowed — and the description, which states what the skill does and when to use it), then the instructions. That simplicity is what makes Skills accessible to business teams, not just developers.

Why this matters for cost: progressive disclosure is also an efficiency lever — you only pay for a skill's context when it is used. This is consistent with token cost governance, covered in the companion article "Governing the cost of AI tokens in the enterprise".

How do you design a good Skill? (the method in 5 steps)

A good Skill starts from a recurring, clearly bounded use case, encodes precise instructions with concrete examples, and is tested before it is shared. A Skill's quality lies in the sharpness of what it describes, not in its length. Here is the sequence we apply and pass on.

1. Choose the right candidate. Aim for a recurring task with a stable expected output, where consistency matters more than pure creativity: writing a scoping note to your template, qualifying a level-1 support request, formatting a report, checking that a document meets your standards. The more often the task recurs and the more identifiable the "right way" is, the better the candidate.
2. Write the description with care. It is the description that decides when Claude triggers the skill. It must state what the skill does and in which cases to use it. A vague description means a skill that never triggers (or triggers far too often).
3. Write clear, sequenced instructions. Describe the method as you would to a new joiner: steps, standards, in-house vocabulary, edge cases, what not to do. Prefer precise instructions to generalities.
4. Give examples. One or two concrete input-to-output examples are worth ten lines of abstract guidance. That is what anchors the "right" result.
5. Test, adjust, then freeze a version. Run the skill on real cases, correct the description and the instructions until the output is reliable, and only then share it. An untested skill that spreads propagates its flaws at scale.

Generic example (client side) — A "scoping note" skill: description = "writes a project scoping note to the internal template; use when the user starts a new project"; instructions = the template, the expected tone, the mandatory sections, the pitfalls to avoid; examples = two model notes. Result: the whole team produces consistent notes, without depending on the one person who "knows how". (Illustrative example — your real skills encode your business, not ours.)

An honest caveat — Not every use deserves a Skill. A rare task, never twice the same, or one that rests mainly on human judgement gains nothing from being fixed — a badly framed skill adds maintenance without value. The rule: you "skill" what is recurring and stabilisable, not everything that works once.

How do you share a Skill across the whole organisation?

On Claude Team and Claude Enterprise, an administrator can deploy a Skill centrally for the whole organisation, or restrict it to a specific group — instead of each person installing it on their own. That is what takes a skill from individual trick to company tool. Central management of Skills at organisation scale was rolled out by Anthropic in late 2025, on those team and enterprise plans.

In practice, several sharing models coexist depending on the environment:

• Org-wide (Team / Enterprise). An organisation owner provisions a skill from the admin settings (Organization settings > Skills); it becomes available to everyone, with no upload needed from each person. Members see it marked as provisioned by the organisation, and can disable it individually if they do not use it.
• By group. A skill can be restricted to a specific team (via a plugin assigned to a group): only the relevant members have access. Useful for skills specific to one business function (legal, finance, support and so on).
• Technical prerequisite. Skills require code execution: the administrator must enable it, along with Skills, in the organisation settings before any deployment.

An important nuance (worth knowing before promising scale): on claude.ai in individual use, personal Skills remain specific to each user and are not managed centrally. Org-wide sharing runs through the Team / Enterprise plans. And by default, the ecosystem provides no native approval workflow for org-wide sharing — hence the need for governance that you put in place (next section).

How do you govern a skills library at scale?

Governing a skills library means putting in place versioning, review before release, consistent naming, security and maintenance — so the library stays a reliable asset instead of becoming an unmanageable mess. This is precisely the work that separates "a few nice skills" from an estate of AI know-how the whole organisation can use.

Five pillars of governance, to install with your teams:

The right governance reflex: start small and clean — a few high-value skills, well named, reviewed, with an owner — rather than a sprawling library nobody maintains. It is the library's quality, not its volume, that creates autonomy.

What are the concrete benefits for the organisation?

Done well, Skills deliver four benefits: consistent quality, faster onboarding, less reinvention, and greater autonomy for your teams. These are not abstract promises — they are the mechanical effects of capturing know-how.

• Consistent quality. Everyone produces to the same standard, whatever their level of ease with AI. The "right way" sits in the skill, not in one expert's head.
• Faster onboarding. A new joiner inherits the team's know-how immediately: skills act as an integration guide, applied automatically.
• Less reinvention. Nobody starts from scratch to find the right prompt again: the work is done once and reused everywhere.
• Greater autonomy. Teams no longer depend on a supplier — or on a single person — to evolve their practices. This is where training shows its full value: we train your teams to create and maintain their own skills, so they become autonomous. (METASENSE runs a real training and skills-transfer activity; it is not certified.)

Worth remembering — A well-designed Skill is know-how that does not leave with the person who held it. Very concretely, it is company knowledge made durable and shareable.

Why bring in support to build your Skills?

Because the difficulty is not writing a SKILL.md file, but choosing the right cases, writing instructions that hold up, and installing governance that keeps the library from drifting. That is where an organisation working alone loses time — and where structured support pays for itself.

Our standing is concrete: we use Claude and its Skills ourselves, every day, to produce work (development, agents, automation, asset generation). So we can tell the skill that holds up in production from the idea that impresses in a demo. It is the Creative Tech double role applied to AI: advisory (what to capture, how to govern) and technical expertise (genuinely understanding how a skill triggers, how it is tested, how it is secured).

And we have done it in real skills transfer:

• Grand Angoulême — conference and agentic AI workshop. A format designed to help non-specialist participants understand and co-build AI use cases — exactly the "we bring the method, you bring the business" logic.
• Bertrandt — HR innovation. An ideation workshop and the design of an AI assistant for CV analysis: precise business know-how, co-built with the teams concerned.
• Strate (design school) & Gocad Lab. Training and Design Thinking modules (Gocad Lab: 4 workshops, 24 staff) — our habit of transferring skills towards autonomy.

Honesty — We do not disclose our own internal skills: they are our core value. What we pass on is a repeatable method so that your teams build their skills, on your business. Business knowledge stays yours; everything is done with your experts.

For a concrete case of AI applied to the web, also read "Do you need an AI agent on your website?".

Where do you start, concretely?

Start with a single high-value skill, on a recurring and clearly identified task — prove it, then extend. Autonomy is built through proof, not through mass deployment. A realistic trajectory:

1. Spot a recurring task where consistency counts (a standardised deliverable, a triage, a check).
2. Write a first skill with the person who "already knows how": their method becomes the skill.
3. Test it on real cases and adjust it until the output is reliable.
4. Share it with the team (org-wide or by group), with a clear name and description.
5. Set minimal governance from that first skill: an owner, a version, a review.
6. Measure, capture, extend to the next skill — and build the library step by step.

Worth remembering — The classic trap: trying to "skill everything" at once. The right sequence: one skill that holds up, one internal proof, governance set from the start, then extension. A library is cultivated, not dumped.

Turn your AI know-how into Skills

Your best uses of Claude deserve better than staying in a handful of people's heads. METASENSE (Vélizy-Villacoublay) helps you choose the right cases, design Skills that hold up, install the governance of your library and train your teams towards autonomy — with Claude as the lead tool, and a method adaptable to other assistants.

Discover our AI & Claude workflow support →

FAQ

What is a Claude Skill?

A Claude Skill is repeatable know-how packaged in a folder (instructions, standards, examples, sometimes scripts) that you teach Claude and that it applies automatically when the task calls for it. It is an official Anthropic product capability, published as an open standard in October 2025. Unlike a one-off prompt, a Skill is reusable and shareable.

How do you create a Claude Skill?

You create a Skill in five steps: choose a recurring, stabilisable task, write a precise description (what the skill does and when to use it), write clear instructions as you would for a new joiner, add one or two concrete examples, then test it on real cases before freezing and sharing it. Technically, a Skill is a folder containing a SKILL.md file.

What is the difference between a Skill and a prompt?

A prompt is a one-off instruction valid for a single conversation, neither captured nor shared. A Skill is reusable know-how, written once, which Claude triggers automatically in the relevant conversations and which can be rolled out across the whole organisation. The prompt is a personal, volatile asset; the Skill is a durable company asset.

How do you share a Skill across the whole organisation?

On Claude Team and Claude Enterprise, an administrator deploys a Skill centrally from the organisation settings: it becomes available to everyone without individual installation. A Skill can also be restricted to a specific group. Prerequisite: enable code execution. On claude.ai in individual use, Skills remain specific to each user.

What is a skills library?

A skills library is an organisation's full set of Skills, organised, versioned and governed — its estate of reusable AI know-how. Kept well (review before release, versioning, clear naming, security, maintenance), it makes teams autonomous and consistent. Kept badly, it becomes an unmanageable mess: quality beats volume.

Do you need to be a developer to create a Skill?

No. A Skill rests on a plain-text SKILL.md file: a description and instructions written like an integration guide. Someone from the business who can explain how they work can design a good skill. Scripts are optional and serve advanced cases. It is precisely this accessibility that lets business teams capture their own know-how.

Does every use deserve a Skill?

No, and that matters. A Skill is justified for a recurring task with a stable expected output, where consistency counts. A rare task, never twice the same, or one resting mainly on human judgement gains nothing from being fixed — a badly framed skill adds maintenance without value. You "skill" what is recurring and stabilisable, not everything.

How do you govern a skills library in a company?

Through five pillars: a review of instructions and scripts before release (org-wide sharing has no native approval), versioning with a named owner, accurate naming and descriptions, security (use trusted sources only, audit the scripts), and periodic maintenance to retire obsolete skills. Better to start small and clean than large and neglected.

Do Skills create a security risk?

A Skill can contain executable scripts: Anthropic recommends using only Skills from trusted sources and auditing their content, because a malicious Skill could execute code unrelated to its stated purpose. The right reflex is to treat adding a Skill like installing software: review before release, especially for sensitive uses.

What business benefit should you expect from Skills?

Four mechanical benefits: consistent quality (everyone produces to the same standard), faster onboarding (a new joiner inherits the team's know-how), less reinvention (the work is done once and reused everywhere) and greater autonomy (teams evolve their practices without depending on one person or a supplier). In short: company knowledge made durable.

Sources

1. Anthropic — Agent Skills (overview, structure SKILL.md, progressive disclosure): platform.claude.com
2. Anthropic — Equipping agents for the real world with Agent Skills: anthropic.com
3. Anthropic — Introducing Agent Skills (open standard): claude.com/blog/skills
4. Anthropic (Help Center) — Provision and manage skills for your organization: support.claude.com
5. Anthropic — Skills for enterprise (deployment & governance): platform.claude.com
6. Anthropic — Public repository for Agent Skills: github.com/anthropics/skills