A Curriculum for AI Agents That Actually Know Go

What it does

This repository packages Go-specific expertise—covering code style, error handling, concurrency, testing, security, and popular libraries like Cobra, Viper, and the author’s own samber/* suite—into atomic skill files. AI assistants load them on demand, so they get relevant rules only when needed rather than carrying a bloated style guide in every prompt. The author bootstrapped the content with Claude Code, then edited, tested, and reworked it by hand.

The interesting bit

The project treats an AI’s context window like scarce real estate: skills are cross-referencing units that mention dependencies on other skills (error-handling conventions live in golang-error-handling, not golang-observability), and the README is refreshingly blunt that AI-generated skills are “useless” without human review. It also publishes per-skill error-rate gaps—some skills cut mistakes by over 60%—based on an evaluation suite of nearly 3,400 test cases.

Key highlights

• Covers 25+ general-purpose skills and 15+ framework-specific ones, from golang-modernize to golang-grpc.
• Works with Claude Code, Cursor, Copilot, Gemini CLI, Codex, and several other agentic coding tools.
• Human-reviewed content: explicitly not raw AI output.
• Evaluation claims 98% pass rate with skills versus 56% without, a +41 percentage point delta across 3,395 test cases.
• Skills are overridable and some trigger “ultrathink” mode automatically.

Caveats

• The README warns that installing only a subset of skills yields a “partial and potentially inconsistent view” because units cross-reference each other.
• One skill (golang-temporal) is marked as to-do with zero content.
• Token counts per skill vary wildly; the heaviest directory (golang-benchmark) weighs in at over 29k tokens, which could still nudge context limits on large prompts.

Verdict

Worth a look if you use an AI assistant for Go and are tired of generic, hallucinated idioms. Skip it if you do not use agentic coding tools or prefer to maintain your own private prompt library.