Playing With Claude Skills in Practice

The stream focused on hands-on experimentation with Claude Skills rather than theory. Kewin Wereszczynski and Jan Jaworski explored how skills work as structured, reusable instructions that guide AI agents when working inside a real codebase. Much of the discussion centered on how skills differ from ad-hoc prompting and how they help constrain agent behavior in practical developer workflows.

AI Agents, Context, and Trust

A recurring theme was context management. The speakers discussed how large context windows can quickly fill up, how that affects output quality, and why breaking work into smaller tasks often leads to better results. They also shared real examples of AI agents making incorrect assumptions, overconfident suggestions, or performing actions that require close supervision, especially when running commands or modifying project files.

React Native and Expo Experiments

During the stream, Claude Skills were tested against a real React Native project. The session included an attempted Expo SDK upgrade using agent guidance, reviewing how the agent handled dependency updates, configuration changes, and error resolution. The experiment highlighted both the speed benefits of agent-driven upgrades and the limits that still require developer judgment, especially around native build errors and platform-specific tooling.

Reviewing Code With Best Practices Skills

The conversation also covered using skills as a review mechanism. A React Native best practices skill was applied to a small codebase to surface issues such as lint warnings, accessibility gaps, debug leftovers, and architectural concerns. This demonstrated how skills can act as a lightweight reviewer rather than a code generator.

Creating Skills and “Human” Output

Toward the end of the stream, the speakers explored skill creation itself. Using a skill-creation workflow, they experimented with generating a custom skill designed to produce more human-like writing. This led to a broader discussion about tone, imperfection, and how AI tends to reveal patterns in human communication when asked to imitate it.