Skills and plugins: package your expertise for Claude

0:00 Skills and plugins

You have a way of doing things. A reporting method, a creative review checklist, a proposal format. Right now it lives in your head or in a document nobody reads. Skills let you package that expertise so Claude applies it consistently, for you and your whole team. In this lecture you will learn what skills are, how to write one, and how plugins bundle them for teams.

0:29 Why package expertise

Why package your expertise at all? Because knowledge that lives in one person's head does not scale. Every new colleague asks the same questions, and every answer is slightly different. A skill turns your method into something repeatable. Think of it like a recipe card in a professional kitchen. The head chef writes it once, every cook follows it, and when the recipe improves, you update one card instead of retraining the whole team.

1:01 Why Skills exist

Project instructions are great for rules. Always use approved prices. Never invent testimonials. But some work needs a procedure, a sequence of steps, reference files, maybe a script. Pasting that into every chat wastes context and produces inconsistent results. Skills solve that.

1:19 Progressive disclosure

A skill is a folder of instructions, and optionally scripts and resources, that Claude loads only when a task needs it. At the start, Claude sees just each skill's name and short description. When your request matches, it reads the full instructions and can run bundled scripts where code execution is available. This is called progressive disclosure, and it means you can install many skills without cluttering every conversation. Skills work in the Claude apps, in Claude Code and through the API, and Anthropic ships built in ones for making Word, Excel, PowerPoint and PDF files.

2:01 Anatomy of a skill

Every skill has a file called SKILL dot M D. At the top sit two fields, a name and a description. Below that, a numbered process, a never list and an output format. Supporting files sit alongside, such as brand rules or approved claims. For a creative review skill, the process might be, describe what you see without judgement, score clarity, brand fit, hierarchy and mobile legibility, quote the rule behind each deduction, check claims against the approved list, check disclosure labels, and output a table with a prioritised fix list.

2:41 Descriptions decide everything

The description is what Claude uses to decide when to load the skill, so it matters enormously. Say what it does and when to use it, using the words people actually type, such as review, score or Q A. Include the brand or client name if it is specific. And keep one job per skill. A giant marketing skill will load at the wrong moments and help nobody.

3:11 Plugins for teams

A plugin bundles skills, connectors and sub agents so an entire workflow installs in one step. In twenty twenty six Anthropic launched a plugin directory, and in September a developer portal for submitting plugins. Vendors are shipping plugins too, for example a Salesforce plugin in beta with dozens of pre built sales skills. On business plans, admins decide which skills and plugins are available.

3:39 Skills can contain code

Remember that skills can contain scripts Claude may run. Treat third party skills like software. Install only from trusted sources, read the SKILL dot M D and any scripts first, watch for instructions that fetch from unknown addresses or send data out, and keep secrets out of skill folders. Anthropic added security scanning for third party skills and plugins, in beta from August twenty twenty six, but reading before installing is still your job.

4:11 Simple example

A simple example. Write a tiny skill called meeting notes. Description, turn raw meeting notes into decisions, actions with owners and dates, and open questions, use when asked to tidy or summarise notes. Process, three numbered steps. Output, three headings. Upload it. Now when anyone types, tidy these notes, and pastes a messy page, Claude loads the skill and produces the same clean format every time. Five minutes of writing, and your team's notes become consistent forever.

4:44 Worked example: a Riyadh agency

A performance agency in Riyadh wrote a monthly client report skill. It holds the steps for pulling the export, their exact metric definitions, a bilingual report template, a script that builds the chart, and a checks needed rule. Any account manager can now type, monthly report for Client B from this export, and get the house format. When a metric definition changed, they edited one file instead of retraining twelve people. When the agency later adopted a plugin that bundled their reporting skill with a connector to their ad platforms, they reviewed its scripts and permissions before enabling it for the team, exactly as they would review new software. The plugin passed the security scan, but they still restricted the connector to read only. Speed and caution, together.

5:40 Try this now

Try this now. Think of one procedure you explain to colleagues again and again, a report format, a review checklist, a proposal structure. Write a SKILL dot M D with a precise name, a description that says what it does and when to use it, a numbered process, a never list and an output format. Add one supporting file, such as your rules or a template. Upload it, then test with three prompts, one that should clearly trigger it, one borderline, and one that should not. Adjust the description until it triggers exactly when you want.

6:22 Watch me do it, part 1

Let me build the creative review skill. I create a folder called creative review with three files. SKILL dot M D, brand rules and approved claims. At the top of SKILL dot M D I write two fields. The name, creative review. The description, review ad creatives, social posts and landing page screenshots against Noor Organics brand guidelines, use when asked to review, score or Q A marketing visuals or copy for Noor Organics. Then the process in six numbered steps. Describe first, score four criteria, quote the rule for each deduction, check claims against the approved list, check disclosure, and output a table with a fix list. Then the never list. No invented predictions, no medical claims.

7:13 Watch me do it, part 2

I zip the folder and upload it in the skills settings, and it appears as enabled. Now the three tests. Prompt one, Q A these four ads, with images attached. Claude shows that it is using the creative review skill and returns the house format table. Prompt two, a borderline one, what do you think of this image? It does not trigger, and I want it to, because people phrase feedback requests that way. So I edit the description to include give feedback on. Prompt three, write me a poem about autumn, correctly does not trigger. After the edit, the borderline prompt triggers too. The skill is ready to share.

8:01 Recap and next step

Recap. Skills package procedures that Claude loads only when relevant, across the apps, Claude Code and the API. Write sharp descriptions, keep one job per skill, and put supporting files alongside. Plugins bundle skills, connectors and agents for teams. And read anything third party before installing. Your next step: pick one procedure you keep explaining to colleagues, write its SKILL dot M D, and test it with three prompts, one that should trigger it, one borderline and one that should not.