Claude skills
How a skill works on the inside.
Always
Discovery
Claude reads only the name and the description of each skill.
If the task fits
Activation
It loads the full instructions in SKILL.md.
Only when needed
Execution
It consults the process's templates, references and scripts.
proposal-review/
├─ SKILL.md name, description, steps
├─ template.md reference
└─ scripts/ optionalClaude skills turn the way your company does a task into instructions Claude applies the same way for the whole team. Today AI use tends to be scattered: everyone has their own prompts and shortcuts, so the improvement stays individual. If you lead HR, L&D or operations, the decision is not a technical one: it is choosing which processes deserve a skill, and who maintains it.
What are Claude skills?
Claude skills are folders of instructions, scripts and resources that Claude loads automatically when a task needs them. Each skill contains a SKILL.md file with a name, a description of when to use it, and the steps of the process. Anthropic introduced them in October 2025 under the name Agent Skills.
The format is an open standard already supported by GitHub Copilot, VS Code, Cursor, ChatGPT and Codex or Gemini CLI. So a well-written skill is not tied to a single tool.
How is a skill different from a prompt or a connector?
A prompt is written every time you need it, a skill is written once and Claude activates it on its own, and a connector (MCP) gives access to a tool without explaining how to use it. The skill supplies the «how»: the criteria, the format and the steps your team already follows.
The technical difference is progressive loading, which works in 3 levels.
| Prompt | Skill | Connector (MCP) | |
|---|---|---|---|
| What it is | A one-off instruction | A packaged, reusable process | Access to a tool or to some data |
| When it acts | When someone writes it | When the task matches its description | When Claude needs to read or write in that tool |
| Who maintains it | Each person | One owner of the process | The technical team or the vendor |
| Example | «Summarise this report» | Reviewing proposals against the company template | Reading the CRM or the ticketing system |
Claude skills · Progressive loading
Claude reads only what the task needs, so your team can keep many skills switched on.
Always visible
Name and description
Claude knows which skills exist and when to use each one.
On demand
The full SKILL.md
Instructions, criteria and the steps of the process.
Only when needed
Reference files and scripts
Templates, examples, rules and code that runs.
Example
proposal-review/
├─ SKILL.md
│ name + description ← 1
│ review steps ← 2
├─ template.md ← 3
└─ scripts/ ← 3The further down, the more detail. That is why SKILL.md stays short and the detail lives in reference files.
Source: Anthropic, Equipping agents for the real world with Agent Skills (2025).
A skill guides Claude, it does not program it. Claude interprets the instructions, so the result still needs human review on critical tasks.
Where can Claude skills be used, and on which plans?
Claude skills work in the Claude apps, in Claude Code and in the API, and are available on the Free, Pro, Max, Team and Enterprise plans, as long as code execution is enabled, according to the help centre. Each person turns them on or off from their own account settings.
Since December 2025, on the Team and Enterprise plans an admin can distribute skills to the whole organisation, enabled by default for everyone. Anthropic also published a directory of third-party skills, from Atlassian, Canva, Figma or Zapier, that can be enabled without building anything.
Check the plans before deciding, because Anthropic updates these features often.
Which tasks are worth turning into Claude skills?
The ones worth turning into Claude skills are the repetitive tasks that already have a working process and someone who owns it. A skill does not fix a confused process: it repeats it faster.
To choose, combine 4 criteria:
- Frequency: the task repeats every week or every month.
- A stable process: the steps and the format are already agreed.
- An owner who knows the process and can review the result.
- A measurable result, in hours or in errors avoided.
The typical cases are recurring reports in the same format, sales proposals that follow a template, reviewing copy against a brand guide, or preparing client meetings.
If you are not sure where to start, an AI consulting engagement maps real use task by task and prioritises 3 to 5 quick wins with estimated hours and euros. There are two formats: Lite (3 weeks) and Full (6 to 8 weeks).
How to write a skill that works in a team
Anthropic's best practices suggest building the skill from where Claude fails, not from everything your team knows. Follow these 5 steps:
- Do the task without a skill. Run it with Claude a couple of times and note what you always correct: format, criteria or missing data.
- Write 3 test scenarios before you write the skill. Real cases with the expected result, so you can compare before and after.
- Write the description around when it triggers. It should say what the skill does and when to use it, in the third person and within 1,024 characters.
- Keep SKILL.md under 500 lines. The detail (templates, examples, rules) goes into reference files Claude consults only when it needs them.
- Assign an owner and a review date. Processes change, and a skill nobody maintains ends up applying old criteria.
---
name: sales-proposal-review
description: Reviews sales proposals before they are sent to the client. Checks structure, tone and mandatory clauses. Use it when someone asks to review, validate or improve a sales proposal.
---
# Sales proposal review
1. Check that the proposal follows the structure in template.md.
2. Verify that no price appears without Finance approval.
3. Return a list of changes ordered by priority.- On
sales-proposal-reviewCustom
Reviews sales proposals before they are sent to the client.
- On
Canva
From the third-party skills directory.
- Off
Figma
From the third-party skills directory.
- Off
Zapier
From the third-party skills directory.
An illustration, not a screenshot. On the Team and Enterprise plans an admin can distribute skills to the whole organisation already switched on.
Test the skill with the models your team actually uses. What works with an advanced model may need more detail with a lighter one.
How are skills governed across an organisation?
Skills are governed with an approved list distributed by the admin, with permissions over who can create and share them, and with one clear rule: only skills from trusted sources get installed. On the Team and Enterprise plans, the admin controls code execution, skill creation, sharing and distribution to the whole organisation.
A skill can include scripts, so it carries the same risks as any third-party software. Anthropic names prompt injection and data exfiltration explicitly, and recommends reviewing the code and the external connections before installing skills.
Claude skills · Governance
Treat every skill as one more tool: approved, with permissions and with an owner.
An approved list of skills
The admin distributes the validated ones to the whole organisation.
Creation and sharing permissions
Decide who can create, share and publish skills.
Trusted sources only
Review scripts and external connections before installing.
An owner and a review date per skill
Processes change: every skill needs someone who owns it.
Ironhack for Business · Based on Claude's own documentation on skills
In practice, add approved skills to the same policy that already defines your authorised tools. If you also need evidence of AI training, our Article 4 test shows which documentation you are missing.
How do you know whether a skill really saves time?
Measure the task, not the skill: how many hours it takes before and how many 60 to 90 days later, valued at the loaded hourly cost. How often a skill triggers tells you about use, not impact.
Give each skill an owner and an indicator: hours per report, proposals reviewed per week, or errors caught before sending. With that number, you decide whether the skill stays, improves or is retired.
For a first figure before you start, the savings calculator estimates the hours a team recovers per year by automating tasks and what they are worth, with editable assumptions. It is an indicative estimate, not a result.
Train your team to write their own skills
The skills that pay off are written by the people who know the process, so the training has to be hands-on and on your own use cases. At Ironhack for Business, our AI workshops are half-day sessions with Claude as the main tool, or with Copilot and ChatGPT depending on what your company already uses.
AI workshops
A hands-on session on your use cases and the tools your company already has.
A 10-minute pre-session survey
Tailors the session to how your team actually works.
Practice on your own use case
A framework, guided practice and a prompting method.
Commitments signed in the room
Everyone commits to applying what they learned.
Material your team keeps
Editable slides, prompt templates, exercises and a replicable demo agent.
- Half a day
- Claude, Copilot or ChatGPT
- FUNDAE-fundable*
*Subject to your company's available credit, which it administers. Licences not included.
The demo agent is a demo, not a deployment into your systems.
Next step
Pick 2 or 3 candidate tasks using the criteria in this guide and turn them into your team's first skills.