24 essentials in your Claude setup

Hey, it's Cindy. You do not need all 24. This is a shortlist organised around useful jobs, with the links and setup distinctions that stop an install list becoming a confusing pile of tools.

Know what you are adding

A skill is reusable instructions, sometimes with supporting files. A plugin bundles capabilities such as skills, commands and tools. An MCP server connects the assistant to external tools or data. Installing instructions does not grant access to an app.

The source packages may be public, but Claude access and connected services can require paid plans, API credits or accounts. Check the README for your exact app before installing. In particular, local Claude Desktop instructions are not automatically compatible with browser chat.

Eight plugins for knowledge work

These come from Anthropic's knowledge-work repository, whose setup instructions cover Cowork and Claude Code. Start with sample files before connecting business accounts.

1. Productivity: turn a messy task list into something you can review. Supply owners and deadlines where known; check that missing details remain visible. Useful first input: your real notes from the week.
2. Sales: use call notes to draft follow-up material and next steps. Compare the draft with the actual conversation before sending. Start with one call, not your whole pipeline.
3. Customer Support: draft a response from a support request and approved policy. Test an exception case so you can see whether it invents a refund or escalates appropriately.
4. Operations: turn a recurring process into a runbook. Give it the trigger, people involved and common failure. Ask a teammate to follow the draft and flag missing steps.
5. Data: explore a spreadsheet and check an analysis. Begin with a small CSV whose totals you know. Verify units, missing values and calculations before trusting the chart.
6. Product Management: synthesise customer interviews into a product brief. Keep quotes attached to sources and distinguish a repeated theme from one person's suggestion.
7. Design: review a screen's hierarchy, labels and accessibility. Provide the intended user and task. Ask for specific issues with proposed fixes, then inspect each recommendation yourself.
8. Enterprise Search: research across connected work sources. Connect only the sources the task needs, then open the cited documents to check whether the conclusion is supported.

Four individual skills for things you make

These are folders in Anthropic's skills repository. For Claude Code, keep each chosen folder and its supporting files together under your documented skills location. Dependencies still matter.

9. internal-comms: draft an internal update from notes. Give it the audience and an approved example; test whether actions and unknowns survive the rewrite.
10. theme-factory: apply a font-and-colour theme to an artifact. Start with one report or presentation and check readability, not only whether the palette looks attractive.
11. slack-gif-creator: build a small animated GIF. Read the helper requirements, generate a short sample and inspect dimensions, file size and legibility in the destination.
12. webapp-testing: inspect a local web app with Playwright. It needs the relevant browser and runtime dependencies. Start with one page and one interaction, then inspect the captured screenshot.

Eight connections for research and app work

Use each provider's linked setup, including its authentication instructions. API keys belong in the documented secret configuration, never a public repository or screenshot. Pick one research provider first.

13. Tavily MCP: search and extract web material. Test a specific research question and open the returned sources. Check the service's current credits and account requirements before larger crawls.
14. Firecrawl MCP: extract structured information from websites. Test a single permitted page with two fields before expanding. Its documentation includes supported Claude setups; service usage may be charged separately.
15. Exa MCP: find relevant pages or company information. Give a narrow topic and compare returned sources. Keep this as an alternative to other search tools until you identify a separate need.
16. Hugging Face MCP: find models, datasets, papers and demos. Follow its supported-client setup and inspect a model card's licence and requirements before treating a search result as a recommendation.
17. Canva MCP: work with Canva designs through the documented design connection. For an individual account, follow the Help Center setup linked from that page and test on a draft. The developer redirect-URL registration instructions are for building an integration, not connecting your personal assistant.
18. ElevenLabs MCP: turn a script into voice audio, or build and test a voice agent, from Claude. It uses ElevenLabs' hosted MCP server (OAuth, nothing to install); the old local repo is archived. Test a short sample before using longer paid generations.
19. Atlassian Rovo MCP: bring Jira and Confluence information into a task. Authenticate the intended workspace and start with reading one issue. Review permissions before allowing record changes.
20. Sentry MCP: investigate an app error using its issue context. Connect the appropriate project, inspect one issue and compare the explanation with the actual stack trace before changing code.

Four Claude Code plugins for building

Browse /plugin and its Discover section, or follow the official directory. Read the individual README before running a command.

21. feature-dev: work through a feature with exploration and review. Bring a small specification and a success check; inspect the proposed scope before implementation.
22. code-review: review a pull request for issues worth investigating. Check its current posting behaviour first. A generated finding is something to verify, not an automatic reason to change code.
23. hookify: turn a rule into a hook. Start with a harmless warning and test both a matching and non-matching case before relying on it.
24. security-guidance: add guidance around risky coding patterns. Check runtime prerequisites and inspect findings. It complements your checks rather than certifying an application secure.

Install one addition and prove it helps

For the Data plugin in Claude Code, open your terminal with Claude Code installed and authenticated, then run the repository's marketplace pattern with the Data package:

claude plugin marketplace add anthropics/knowledge-work-plugins
claude plugin install data@knowledge-work-plugins

Open a new Claude Code session in your project folder, check /plugin for the installed entry and save this as weekly.csv:

channel,leads
Search,12
Social,8
Referral,5

Ask the Data plugin to calculate total leads and each channel's share, showing its calculation. The total should be 25 and the shares 48%, 32% and 20%. Those are practice data, not business results. If it requests a warehouse connection, explain that the input is the local CSV.

For Cowork, use the plugin directory linked in the repository's Getting Started section and choose Data. For an individual Code skill, download its whole folder, inspect it and place it at .claude/skills/internal-comms/ with SKILL.md directly inside. Keep its referenced files alongside it. Restart the session and request an internal update using that skill.

MCP setup is different: a remote server uses a service URL and authentication, while a local server starts a process on your computer and needs its runtime installed. Use the exact provider instructions for your Claude surface. Pasting a remote URL into a local command field will not configure the connection.

Check the output against the source. If it asks for a connection your task does not need, use local input first. If installation fails, check the app surface, package name and prerequisites before installing alternatives. If the result is weak, improve the input and instructions before adding more tools.

My next task is [task], and I use [Claude surface and plan]. From this shortlist, choose the smallest addition that directly helps. Read its current official setup instructions and explain whether it is a plugin, skill or MCP connection, what it can access, and any separate account or usage costs.

Walk me through installation without guessing commands. Then help me test it on [small input] against [known result or checklist]. Show the output, identify unsupported assumptions and tell me how to disable the addition if it does not help. Do not connect extra accounts, publish anything or send messages as part of this test.

Keep the addition when it improves a real task you understand. That is a more useful setup than a long installed list. Explore all free guides when you have your next job in mind.