Microsoft dropped a curious open-source project on Hacker News today: AI Engineer Coach, a VS Code extension that promises to give developers the same kind of granular visibility into their AI coding habits that fitness apps provide for workouts. The pitch is straightforward—track your progress, detect anti-patterns, and level up how you work with tools like Copilot, Claude, and Codex.

How It Works

The extension reads your local AI session logs and performs all analytics entirely on-device. Nothing gets shipped to external servers, which should ease enterprise concerns about sensitive code ending up in training pipelines. The project makes a point of being read-only—it never modifies your session files—and explicitly states it doesn't phone home or collect usage data. At its core, AI Engineer Coach parses your coding assistant interactions and transforms them into structured insights across multiple dashboards. You can filter by workspace or harness (the specific tool you're using), making it relatively easy to compare how you work with different models or in different contexts.

Feature Breakdown

The tool is organized into three main sections: Observe, Measure, and Improve. The Observe tab gives you practice scores with week-over-week trends, a daily activity chart, and workspace-level statistics. There's also a Gantt-style timeline view for drilling down into individual sessions and detecting overlaps in your work patterns. On the measurement side, you're looking at generated code volume broken down by programming language, plus model usage tables (though token breakdowns are currently disabled). There's also an activity heatmap showing 7×24 patterns and what the docs call "work-life balance signals"—an interesting framing that suggests this tool is thinking about sustainable practices rather than just raw productivity.

Anti-Pattern Detection Is Where It Gets Interesting

The Improve section houses 45 detection rules spread across five practice score cards: prompt quality, session hygiene, code review, tool mastery, and context management. Each rule comes with severity ratings and concrete remediation actions. The Rule Editor lets you create and tune these visually or as raw markdown, with a live-testing feature against your actual data. There's also a Rule Playground—an interactive REPL for the rule DSL that includes a field browser, function catalog, and metric list. For developers who want to understand how the detection system works under the hood, this is where you'd dig in. The Data Explorer gives you direct access to session fields with distributions and ad-hoc filtering capabilities.

Skill Discovery and Context Health

Two features stand out for their forward-thinking approach: Skill Finder discovers repeated prompt patterns and matches them against an open-source community skill catalog—essentially turning your most-used prompts into reusable assets. The Context Health dashboard runs agentic readiness checks, audits instruction files, and generates workspace context maps with AI-powered review (using VS Code's built-in Copilot API when you explicitly invoke it).

Gamification Layer

Microsoft didn't skip the engagement hooks. There's an XP-based achievement system progression through Bronze → Silver → Gold → Diamond tiers, personalized quizzes generated from your actual usage patterns, and a shareable stat card for flexing (or humble-bragging) about your AI engineering habits. The Learning Center generates code-comparison rounds based on how you've actually been using these tools—more useful than generic tutorials if the underlying data is good.

Key Takeaways

  • All processing happens locally—no telemetry, no phone-home behavior
  • Supports any harness (Copilot, Claude, Codex) through session log parsing
  • 45 customizable anti-pattern detection rules with severity ratings
  • Gamification elements target developer engagement and habit formation
  • Currently in early stages (4 points on HN at publication time)

The Bottom Line

This is a genuinely interesting concept—if you've ever wondered whether you're actually getting better at AI-assisted development or just building new dependencies, this gives you the instrumentation to find out. Whether it gains traction depends heavily on whether developers see value in quantifying their AI workflows the same way they track test coverage or deployment frequency.