Claude Skills: A Practical Guide

Claude skills are folders of instructions, built around a SKILL.md file, that tell Claude how to do a specific job. Claude reads each skill's description and uses the skill when a request matches, or you can call it directly. You add a skill by uploading it as a ZIP file in Claude's settings, and team plans can share skills across an organisation.

A Claude skill is a job card for Claude. Like a new starter's job card, it sets out the steps, the format and the checks for one task, and points to the templates and examples the task needs. Claude already knows how to write and analyse; the skill tells it how your business does this particular job.

Each custom skill has a SKILL.md file with two required details: a name of up to 64 characters and a description of up to 200 characters that says what the skill does and when to use it. Claude relies on the description to decide when to use the skill, so it's the most important line you'll write (Anthropic: How to create custom skills). Supporting files, such as a reference document or a script, sit in the same folder.

Anthropic also provides its own skills for creating Excel, Word, PowerPoint and PDF files. Claude uses these automatically when they're switched on (Anthropic: Use skills in Claude).

The format came from Anthropic and is now an open standard, the Agent Skills format, used by many other AI tools (Agent Skills overview). See What Are Agent Skills? for the general picture.

All of the following comes from Anthropic's current guide (Anthropic: Use skills in Claude):

• Plans. Skills are available on Free, Pro, Max, Team and Enterprise plans, with different management options on each.
• Switch them on. Skills need code execution and file creation turned on. On individual plans that's in your settings. On Team and Enterprise plans, an owner enables skills in the organisation settings.
• Install a skill. Package the skill folder as a ZIP file, open Customize > Skills, select the + button, choose Upload a skill and select the file.
• Turn skills on or off. Customize > Skills lists your skills with a toggle for each.
• Share with your team. On Team and Enterprise plans you can share a skill with colleagues, with groups on Enterprise, or publish it to your organisation's library. Recipients can use but not edit a shared skill, and they get your updated version at next use.
• Provision for everyone. Owners can make a skill available to all users in the organisation.

To use a skill, just ask for the work. Claude matches your request against each skill's description and loads the one that fits, and it can combine several skills in one task. Anthropic's plugins guide also describes picking a skill yourself by typing / or using the + button in the message box.

Skills can also arrive inside plugins, which bundle skills, connectors and sub-agents for a role or team (Anthropic: Use plugins in Claude). For how these pieces differ, see Claude Skills vs Plugins, Projects and Commands.

Brightwater Accounting, a fictional practice, sends new clients an engagement proposal. Partners write them differently, fees sometimes drift from the fee schedule, and some proposals skip the scope exclusions.

Their Claude skill:

• Name: engagement-proposal
• Description: Drafts an engagement proposal from discovery notes using the fee schedule. Use for new client proposals and fee letters.
• Inputs: discovery notes, the client's entity type and the current fee schedule.
• Steps: summarise the client's situation, list the services, price each from the fee schedule, state what's excluded and set out next steps.
• Decisions: never quote a fee that isn't on the schedule. If the entity type is missing, ask. Mark any service the client hasn't confirmed as "to confirm".
• Checks: every fee matches the schedule, exclusions are listed, and the next steps name who acts.

The skill doesn't contain the fee schedule itself. It names it, so updating fees doesn't mean editing the skill. Claude still needs access to wherever the fee schedule lives, through a connector or an uploaded file. If it can't reach the document, the skill should say so rather than quote from memory.

• Write or download a skill for one job, with a specific description.
• Put SKILL.md and any supporting files in a folder named after the skill.
• Compress the folder as a ZIP file, with the skill folder at the top level.
• Check that code execution is on, or ask your organisation's owner.
• Open Customize > Skills, select +, choose Upload a skill and pick the ZIP.
• Make sure the skill's toggle is on.
• Ask Claude for the job in your own words, then run the three tests below.

• A normal request. Use real discovery notes. Check each fee against the schedule and each service against the notes.
• Incomplete input. Leave out the entity type. Claude should ask for it rather than assume a company.
• An unrelated request. Ask for a client newsletter. The proposal skill shouldn't load. If it does, make the description narrower.

Keep each skill to one job; Anthropic advises separate skills for different workflows rather than one combined skill (Anthropic: How to create custom skills). Be specific in the description about when it applies. Include a short example when it clarifies the output. Test after each change, and change only the part that failed. When an edit makes a skill worse, you need the previous version to hand.

Claude's sharing works well inside one Claude organisation. Contexta is for teams that use Claude alongside other AI tools and want one maintained set of skills and company context across all of them. A skill tells your AI what to do. Your knowledge base shows it how, in detail. Your company context tells it who it's for. The Contexta harness brings them together, in every AI.

Contexta applies your company context, covering your terminology, naming and formats, under every skill, for the company and for each team. Skills point to the fee schedules, templates and past work in your knowledge base, which stays in your own SharePoint or Google Drive. Company, team and personal spaces give each group its own home and rules. Drafts stay private until published, and version history lets you compare and restore.

Claude is tested with Contexta and connects without code. Connected Claude sessions fetch the published skill on next use; a skill you download and upload yourself is a separate copy. See How It Works and Integrations.