How to Install Skills in Claude Cowork and Claude Code (Step-by-Step)
If you’ve been anywhere near Claude Code or Claude Cowork lately, you’ve probably heard the word “skills” thrown around a lot. Skills are how you extend Claude beyond generic chat and turn it into a specialist for whatever you’re building — a video creation pipeline, a GHL automation builder, a content engine, whatever. The problem is most people either grab a random skill off a marketplace site without checking it, or they try to build one from scratch and give up halfway through.
Want to go deeper? I run a live Claude Cowork workshop — check the latest dates and see if you can still grab a seat.
Below is the exact process for how to install Claude skills the right way — researched first, validated for security, then built (or customized) with Claude’s help — whether you’re working inside Claude Code or Claude Cowork. Watch the full tutorial here to see the whole build live.
What Are Claude Skills, and Why Should You Care?
A skill is a packaged set of instructions, context, and sometimes scripts that Claude loads when it’s relevant to the task at hand. Instead of re-explaining your process every single session, you install the skill once and Claude already knows the docs, the tools, the SOP, and the guardrails.
For an operations-heavy business, that’s the difference between “AI as a chatbot” and “AI as an actual team member.” You can build a skill for GHL snapshot deployment, client onboarding, content repurposing, video generation — anything that follows a repeatable process.
Step 1: Research Before You Build Anything
Before touching a single line of a skill file, research the tool or workflow you’re trying to teach Claude. In this example, the target was building a skill around Remotion and HeyGen’s Hyperframes for AI video generation — two tools neither creator nor Claude had deep expertise in yet.
The process:
- Search the tool name plus “skills” (e.g., “Remotion skills”) to see what’s already out there — GitHub repos, official docs, existing Claude/Claude Code skill packages, and tutorial videos.
- Go straight to the official documentation site first. That’s your source of truth.
- Cross-reference with a second search pass to catch anything the first search missed, including dev-focused walkthroughs and existing community skills.
Why Not Just Download the First Skill You Find?
Because not every GitHub repo or “skills marketplace” listing is trustworthy. A repo with 3.5 stars and no real activity is a different risk profile than an official vendor repo. Star count and install count are useful signals, but they are not verification. Treat every third-party skill or repo as unverified until you’ve actually checked it.
Step 2: Organize Your Research Into a Second Brain
Once you start pulling from docs, GitHub, and tutorials, you need somewhere to dump it all before Claude can make sense of it. This is where a “second brain” setup comes in — a dedicated Obsidian vault connected to a folder that Claude Code can also read from.
Practically, that means:
- Open the folder in your code editor (VS Code, or whatever you use) so Claude Code has direct file access to the same notes you’re collecting in Obsidian.
- Web-clip sources directly into that vault as you research, instead of trying to hold everything in your head or in scattered browser tabs.
- Keep the second brain scoped to what you’re building right now. If it’s big enough to need its own reference system, give it one; otherwise you’re overcomplicating a simple task.
Step 3: Feed Claude the Sources and Let It Catalog Everything
Start a fresh conversation and tell Claude exactly what you’re doing: you want to build a skill (or a set of skills) from scratch, you’re going to feed it sources one at a time, and it should only start designing the skill once everything has been provided.
Then hand it sources one by one — official docs first, then GitHub repos, then any existing community skills you found. Claude can search and pull from these directly if it has web tools connected (Playwright, Perplexity, or similar MCP connectors).
Should You Trust Third-Party Skill Repos?
No, not by default. Before incorporating any external repo into your own skill, have Claude run a security pass on it: check every file for prompt injection attempts, hidden instructions, or anything that could execute unwanted actions. This is a non-negotiable step, not an optional one. A skill file is still just text that gets loaded into Claude’s context — if it contains hidden instructions, it can hijack the session exactly the way a malicious webpage or email can.
Step 4: Brainstorm the Shape of the Skill Before You Build It
Once Claude confirms it has enough source material, don’t let it jump straight to building. Run a brainstorming pass first. Claude will typically ask you to define:
- Scope — is this one tool, or a combination of tools working together?
- Use case — marketing ads, social clips, course content, client deliverables?
- Audience — is this just for you, for your team, or for clients and students who’ll use it downstream?
- Execution environment — local only, or does it need to scale to cloud rendering/processing?
- Architecture — one mega skill, a hub with multiple sub-skills, or several standalone skills that reference a shared SOP?
Answering these up front keeps you from ending up with a bloated, unfocused skill that tries to do everything and does nothing well.
Step 5: Understand the Real Cost of Installing Multiple Skills
A skill set built this way can easily end up as a dozen-plus individual skills. That sounds heavy, but it’s important to understand what’s actually loaded into context:
- At session start, only each skill’s name and description load — not the full body.
- The full skill content only loads when it’s actually invoked or read.
- Metadata for a large skill set typically eats a small fraction of your total context budget, not a meaningful chunk of it.
That’s why scoping a large skill set to a specific project folder or a dedicated agent (rather than installing it globally for every Claude session) is the smarter move. You get the full toolkit exactly when you need it, without cluttering every unrelated conversation.
How to Install Skills in Claude Code
Once the skill set is built:
- Save the finished skill files to a dedicated folder (a downloads folder works fine as a staging area).
- Move that folder into your dev workspace, inside a project folder scoped to the specific use case — for example, a dedicated “agent” folder rather than your general workspace.
- Open that folder directly in your code editor and start a new Claude Code session from inside it.
- Tell Claude to build (or configure) an agent around the newly added skills, so future requests in that project automatically route to the right skill set.
This keeps skill sets project-scoped instead of globally installed, which matters once you’re running more than one specialized workflow.
How to Install Skills in Claude Cowork
Claude Cowork handles this a little differently since it’s built around connectors and a simpler skill-creation flow:
- Go to Customize → Skills inside Cowork.
- Click the plus (+) icon to create a new skill.
- Feed it the same kind of source material — documentation links, research notes, transcripts — and let Claude help you draft the skill using its built-in skill creator.
- Save and the skill becomes available to invoke in future Cowork sessions.
One limitation worth knowing: a skill (or agent) you build for Claude Code isn’t automatically portable into Claude Cowork. If you want equivalent functionality on both sides, you’ll generally need to either recreate the skill natively in Cowork, or look for an MCP connector that exposes the same functionality through Cowork’s connector browser. Not every niche tool has one yet — check “Browse Connectors” inside Cowork before assuming you need to build from scratch.
Putting It to Work: Building an Agent Around the Skill
Once the skill set is installed at the project level, the real payoff is wiring it into a dedicated agent. In practice, that means telling Claude Code (from inside that project folder) to create an agent that automatically uses the installed skills any time a relevant request comes in — without you having to manually invoke anything.
From there, a request like “create a short explainer video using the info in my second brain” is enough. Claude pulls the relevant source material, applies the skill’s process end-to-end, and delivers the output straight to your downloads folder.
Frequently Asked Questions
Is it safe to download Claude skills from third-party marketplaces?
Not automatically. Treat every third-party skill the same way you’d treat an unfamiliar npm package or browser extension: check the repo, check the files, and have Claude scan for prompt injection or hidden instructions before you incorporate it into anything you use regularly.
Do more installed skills slow Claude down or eat my context window?
Barely, at rest. Only the skill’s name and description load at session start; the full instructions only load when a skill is actually invoked. A large skill set costs you a small percentage of your context budget just sitting there — it’s invocation, not installation, that pulls in the heavier content.
Can I use the same skill in both Claude Code and Claude Cowork?
Not directly. The two run in different environments, so a skill built for one doesn’t automatically transfer to the other. You’ll typically need to rebuild it natively in Cowork, or connect an equivalent MCP connector if one exists for that tool.
Should I install a skill globally or scope it to one project?
Scope it, especially for large or specialized skill sets. Installing a dozen-plus skills globally means they load into every session whether you need them or not. Keeping them tied to a specific project folder or agent keeps things clean and keeps unrelated conversations from getting cluttered.
What’s the fastest way to start building a custom skill?
Research first, build second. Pull official docs, validated GitHub repos, and any existing community skills into one place, have Claude catalog and security-check all of it, then run a quick brainstorming pass on scope and architecture before asking Claude to build anything.