Skills, Cowork & the AI Fluency Framework
An exploration of Claude’s Skills & Plugins system, the three interaction modes (Chat, Cowork, Code), Anthropic’s 4D AI Fluency Framework, and an honest conversation about AI data privacy.
Key Takeaways
- Skills are reusable instructions Claude remembers — but never download unknown skills from LinkedIn, as they can contain hidden API calls that leak your data
- Claude has three modes: Chat (connectors & projects), Cowork (local folder access & browser control), and Code (development environment)
- Anthropic’s 4D Framework guides AI interaction: Delegation → Description → Discernment → Diligence — connecting strategy with tactics
- Always write a diligence statement declaring what AI did versus what you did — transparency builds trust and accountability
- Anthropic pledged no ads and no training on user data, unlike OpenAI which is already testing advertisement in ChatGPT
What We Covered
Skills & Plugins in Claude
Otakar introduced Claude’s extensibility system — Skills and Plugins — which allow you to add custom capabilities beyond the base model:
- Skills – Verbal instructions that Claude remembers and can trigger on command. For example, saying “onboard a new hire” triggers the predefined steps you’ve described
- Plugins – Collections of multiple skills grouped by department or role. An HR Plugin might include onboarding, offboarding, and review skills
- Security warning – Be very careful downloading skills from LinkedIn or other sources. Skills can call external APIs, and hidden code can exfiltrate your data. Always review and get inspired, but build your own
The key distinction: skills should reflect your workflow, not someone else’s. Claude also offers built-in default skills, including one that helps you create your own new skills.
Claude Chat vs Cowork vs Code
Otakar demonstrated the three distinct modes of Claude’s desktop app, each optimised for different types of work:
- Chat – Supports projects, connectors (Gmail, Calendar, Notion, Airtable), and is available on mobile. You can turn connectors on and off per conversation
- Cowork – Accesses a local folder on your computer (like Google Drive synced locally), has browser control, and supports plugins & skills. Not available on mobile. Still in beta
- Code – For coding and development tasks
The main advantage of Cowork: if your data is on your computer, you don’t need to re-upload it to a project — Cowork accesses it directly from the chosen folder. For on-the-go work, Chat with projects is better since Cowork can’t run on your phone.
The AI Fluency Framework (4Ds)
Otakar walked through Anthropic’s structured framework for effective, efficient, ethical, and safe AI interaction:
- Delegation – Define your goal. Understand what AI can do for you and what you are actually good at. Think in the big picture before you start
- Description – Write your prompt. Describe as best as you can what you want AI to achieve
- Discernment – Evaluate the result. Does it move you toward your goal? If not, iterate back to Description — or even back to Delegation if the whole goal needs adjusting
- Diligence – Safety, ethics, fairness, and transparency. Ask: is this result ethical? Is it discriminating anyone? Write a diligence statement
The framework operates on two levels: strategic (Delegation + Diligence) and tactical (Description + Discernment). All four are interconnected — you can loop between any of them.
“Every time you work with any AI tool, you should be able to write a statement saying: I used this AI tool using this process, this was my participation, and this was the participation of the AI tool.”
AI Data Privacy
A wide-ranging discussion about what “be careful with AI” actually means in practice:
- OpenAI’s approach – Already testing advertisements in ChatGPT, which means giving advertisers access to user data for targeting. Data leaks have already occurred
- Anthropic’s stance – Publicly committed to never having advertisements in Claude. Never trains on user conversations
- Personal data – Even with privacy-focused platforms, remain cautious with highly sensitive information. Toggle off “train on my data” settings wherever available
- Practical advice – Give AI specific, smaller tasks that you can easily verify. Don’t give it ambiguous control over your inbox or important data
Use Cases
Topics and challenges participants brought to the session.
Questions Asked
Q What is the difference between a skill and a workflow?
A skill is a verbal explanation to AI — Claude remembers the steps and triggers them on command. A workflow is a broader process, often involving automation tools like Make or Zapier. In practice, a skill can represent a workflow, but the terminology differs: skills live inside AI tools, while workflows describe your overall process.
Q Does the Claude app consume a lot of internet bandwidth when running?
It shouldn’t under normal use. However, Claude Cowork is still in beta and may have bugs that cause higher-than-expected resource usage. You can use bandwidth monitoring apps to check what’s consuming your internet. If the issue persists, try closing Cowork when not actively using it.
Q Can I do the same things in Chat that I do in Cowork?
Yes, Chat supports connectors (Gmail, Notion, Calendar, etc.) and can read and create content across those tools. The key difference is that Cowork also accesses a local folder on your computer and offers browser control. Chat is better for on-the-go work since it runs on mobile, while Cowork is best for project work at your desk.
Homework
- Watch the skills video on Sooper and explore Claude’s built-in skill for creating your own new skills
- Try Claude Cowork with a local folder for one real task this week — start small and verify the results
- Check your AI platform privacy settings — disable “train on my data” in ChatGPT if not already done