From Prompting to Process: What Changed When Flutter Shipped Agent Skills

Google is introducing Flutter’s engineering workflows as machine-readable guidance for AI agents. While this may seem like another AI feature, it signifies a significant shift in how teams build with Flutter. Previously, AI coding tools like Cursor, Claude Code, Copilot, and OpenCode produced code, explained APIs, wrote tests, and navigated large codebases with good accuracy.

However, the consistency of outputs across different sessions or models was often lacking. Teams would resort to writing better prompts, adding repository rules, or creating custom skill.md files to steer AI toward correct decisions.

Google took a similar approach by documenting their architecture, coding conventions, review expectations, and engineering practices as custom Skills. They found that these Skills made AI agents much more consistent. When Flutter announced official Agent Skills, curiosity about their use grew into experiments to understand what these Skills were actually solving.

Flutter’s AI developments were interconnected. Rules helped define project-specific conventions and preferences, MCP allowed AI agents to interact with running Flutter applications, and Agent Skills answered how Flutter recommends solving problems. This shift means Flutter now versioning its engineering workflows alongside the framework itself, providing structured workflows that AI agents can follow.

To test this, experiments were conducted on declarative routing. With Agent Skills, the agent followed Flutter’s own workflow after selecting the appropriate skill, compared to previous inconsistencies. When both Flutter's and custom Skills were available, the agent used Flutter’s Skills for framework guidance and the custom Skills for application-specific workflows.

This integration of Flutter’s engineering knowledge into AI agents removes duplication, streamlines processes, and keeps workflows updated with Flutter’s recommendations.

However, while Flutter’s Agent Skills provide a strong foundation, they do not replace an organization’s own architecture, coding standards, or business-specific workflows. These responsibilities remain with the organization. In essence, Flutter is taking ownership of its engineering expertise, allowing engineering teams to leverage AI without worrying about the specifics of how Flutter expects certain features to be implemented.