Claude Skills 101: with Cowork and Code

Section 1: Claude in 2026 and Where Skills Fit

• The toolset: the Claude app, Claude Code, Projects, connectors, skills and plugins

• The Claude app after Cowork merged into chat: what changed, what stayed and the role of Claude Code

• When to use a skill instead of project instructions, CLAUDE.md, a connector (MCP) or a subagent

• The Agent Skills open standard: write a skill once and use it in many tools

Section 2: Anatomy of a Skill and Progressive Disclosure

• Skill folder structure: SKILL.md (required), references, assets and scripts

• YAML frontmatter: name and description, and the rules the Claude app checks on upload

• Layered loading: Claude sees only the name and description first, then loads the full skill when a task matches

• Writing a good description: what it does, when to use it, and the words users actually say

Section 3: Lab: Designing Skills from Your Team's Real Work

• Analyse repeated work: which tasks make good skills and which are not worth it

• Reference skills (knowledge and standards) versus task skills (step-by-step procedures)

• Keep skill content concise, with input/output examples and the questions to ask the user first

• Lab: move long content into reference files to save context

Section 4: Lab: Creating and Managing Skills in the Claude App

• Turn a conversation that already works well into a skill with skill-creator

• Write a skill yourself, zip it, upload it through Customize and switch it on or off

• Let Claude pick a skill from your request, or call it directly from the / menu

• Combine your skills with Anthropic's ready-made skills for Word, Excel, PowerPoint and PDF

Section 5: Lab: Skills with Files, Connectors and Automation

• Let a skill read several files you allow and produce a document, slides or a spreadsheet

• Connect connectors such as Google Drive, Gmail or Slack and define in the skill what to fetch or send

• Set up a scheduled task so a skill runs on a cycle, such as a report every Monday morning

• Unattended runs: require confirmation before sending messages or changing data

Section 6: Workshop: Ready-made Plugins and the NovaDesk Skill Kit, Day 1

• Add plugins from Anthropic's directory and Knowledge Work Plugins, and review their skills before use

• Build novadesk-brand-voice to set the brand's tone and writing style

• Build weekly-sales-report and meeting-to-actions from sample sales files and meeting notes

• Test the skills working together and schedule the sales report to run every week

Section 7: Lab: Claude Code for Every Learner

• Start Claude Code from the CLI, the Code tab in Claude Desktop or the VS Code extension

• Where skills live: personal (~/.claude/skills/), project (.claude/skills/), plugins and skills synced from your claude.ai account

• Custom commands now merged into skills, and running a skill with /skill-name

• List loaded skills with /skills and see which one wins when names clash

Section 8: Lab: Arguments, Supporting Files and Dynamic Context

• Take user input with $ARGUMENTS, $0, $1 and named arguments

• Add Python or Bash scripts to a skill and reference them with ${CLAUDE_SKILL_DIR}

• Dynamic context injection with the ! syntax to pull live data, such as git diff, before Claude starts

• Claude Code-only features compared with standard fields that work everywhere

Section 9: Controlling Invocation, Permissions and Skill Safety

• disable-model-invocation and user-invocable: decide who can trigger a skill, useful for high-impact tasks

• allowed-tools pre-approves tools and disallowed-tools blocks tools while a skill runs

• context: fork and agent: run a skill in a separate subagent for research or review work

• Read other people's skills and plugins before installing, and never put passwords or API keys in a skill

Section 10: Lab: Testing and Measuring Skills Systematically

• The baseline idea: compare results with and without the skill

• Use skill-creator to build test sets, measure pass rate, tokens and time, and compare two versions

• Tune the description from trigger tests, covering prompts that should and should not use it

• Check each skill's context cost and usage with /skill-doctor

Section 11: Lab: Packaging Skills as a Plugin and Sharing Them

• Plugin structure: .claude-plugin/plugin.json, the skills/ folder and extras such as agents, hooks and MCP

• Create a marketplace with marketplace.json on GitHub and add it with /plugin marketplace add

• Check the structure with claude plugin validate and test the whole plugin with claude plugin eval

• Install the same plugin in the Claude app and Claude Code, and share project skills through Git

Section 12: Master Workshop: NovaDesk Skill Kit and Scaling Up

• Bring the three Day 1 skills into a repository and add a Python check of the sales file to weekly-sales-report

• Build pr-summary, which pulls the diff with dynamic context and runs in a subagent, plus a user-only release-checklist

• Test the main skills with baseline evals and tune their descriptions from the results

• Package everything as the "NovaDesk Skill Kit" plugin, publish it through an internal marketplace and install it on both sides

• Scaling up: naming, skill owners, review cycles and next steps with the Claude API, Agent SDK and MCP