Claude Code: My experience with AI fluency rubrics

Engineering efficiency isn't just about picking the right LLM; it's about how well your team manages context and agentic workflows. I've been looking into how different developers approach tools like Claude Code and Cursor, and the gap between a power user and a novice is massive. It usually comes down to prompt engineering precision and how they handle the agent's workspace.

I recently came across a project called Promptster that attempts to quantify this "AI fluency." Instead of just guessing who is productive, they are building a platform to analyze how engineers actually interact with AI coding agents. They've even released an open-source rubric to define what "best practices" actually look like when working with Claude Code or Copilot.

From an engineering standpoint, the focus on context management is what caught my eye. Most people struggle because they feed the agent too much noise or not enough specific architectural constraints. A good rubric for an AI agent should probably evaluate:

Context Window Discipline: How effectively the dev provides relevant file snippets without bloating the prompt.

Instruction Granularity: Whether the commands are specific enough to prevent the agent from hallucinating logic.

Verification Loops: How the developer validates the agent's output against existing test suites.

If you are interested in how to standardize AI usage across a team or want to see a framework for measuring agentic skill, check out their open-source rubric here:

https://github.com/promptster-ai/rubric

It’s a much more structured way to look at productivity than just tracking lines of code. For those of us focused on performance, seeing a dashboard that tracks DORA metrics alongside AI fluency is a pretty interesting way to bridge the gap between traditional DevOps and the new AI-driven workflow.