[Masterclass for Pro] Part II : Creating Advanced Custom gAI Workflows and Tools for Scientists

Instructeur:

Marin Bogdan, Msc Phyics ETH Zürich

Independent IT consultant specialized in AI, project management and information security
Quality manager and AI lead for Swiss E-Voting, Swiss Post Ltd.​

Topics Covered

Workflow Automation & Skills

The session opens with the big-picture theme tying everything together: turning your AI from a passive question-answering tool into an active collaborator that executes your work. You’ll learn how to teach an AI your own skills and repeatable procedures so it can carry out complete, multi-step workflows on your behalf rather than answering one prompt at a time. This section introduces the vocabulary and building blocks used throughout the rest of the training — Skills, Plugins, MCP Servers, Automation, Remote Work, Context Management, and Markdown — setting the foundation for the more focused topics that follow.

Skills & Plugins

Here you’ll learn to extend your AI assistant with custom expert knowledge and packaged workflows specific to your own research or teaching domain. Rather than re-explaining context every time, Skills and Plugins let you encode specialized procedures once — such as a standard literature-review method, a data-cleaning pipeline, or a grading rubric — so the AI can apply that expertise consistently and automatically whenever the task calls for it.

Artifacts

This segment covers Claude Artifacts, a feature for producing shareable, disposable mini-tools directly inside a conversation. Instead of writing a full application, you generate a small, self-contained tool — a calculator, visualization, or interactive form, for example — that can be used immediately, shared with colleagues, or discarded once its purpose is served. It’s aimed at quick, low-stakes utility creation rather than long-term software development.

MCP Servers

This topic addresses how to connect your AI to local and remote MCP (Model Context Protocol) servers, effectively giving your AI “hands” to complement its “brain.” Where a chat model can only reason and respond in text, an MCP connection lets it actually interact with external systems — reading files, querying databases, calling APIs, or triggering real actions — turning conversational AI into an operational agent within your existing tools and data.

On-the-Fly Tools: Beyond Vibe-Coding

This section teaches you to build small NodeJS-based applications with AI assistance and publish them quickly, even if you don’t have a traditional programming background. It goes a step beyond casual “vibe-coding” (loosely prompting an AI to produce code without real understanding) by introducing a more structured, repeatable approach to rapidly shipping small, functional apps you can actually rely on and maintain.

AI-Assisted Software Engineering

This topic explores how AI-assisted development enables you to write secure, reliable code far faster than manual coding alone, using spec-driven workflows. Rather than prompting ad hoc, you start from a clear specification of what the software needs to do, and the AI helps implement it against that spec — improving reliability, traceability, and code quality compared to unstructured AI-generated code.

Hands-On Examples

The training closes with practical, guided exercises that let you apply the concepts from all previous sections in real scenarios. This is where the skills, plugins, MCP connections, artifacts, and spec-driven coding practices come together in concrete, worked examples rather than abstract explanation.

Technical Requirements and Reccomendations

Attendees must arrive with the following set up in advance:

• A laptop/notebook computer

• A free GitHub account with GitHub Copilot’s free subscription tier

• A Pro or Max subscription to Claude.ai [Optional]

• Claude Desktop and Claude CLI installed [Optional]

• VS Code installed [Reccomended]

• The Claude Code extension for VS Code installed [Reccomended]

• GitHub Copilot CLI installed [Reccomended]

Note: Several requirements and recommendations involve paid subscriptions (Claude Pro/Max) and multiple software installations. Attendees should anticipate setup time beforehand to avoid delays on the training day. Participation is also possible without a subscription, although hands-on personal practice may be limited. While the workshop uses Claude/Anthropic and GitHub Copilot as the primary illustrative examples, it also discusses equivalent alternatives from other vendors.