What it does Learn Harness Engineering is a 12-lecture, 6-project course on building the environment around LLM coding agents. It covers instruction files (AGENTS.md, CLAUDE.md), progress tracking, verification pipelines, scope control, and session lifecycle management. The repo includes a quick-start skill under skills/harness-creator/ that scaffolds a production-grade harness for your own project, plus a PDF build pipeline and GitHub Actions workflow for publishing coursebooks.

The interesting bit The course is built on a sobering finding from Anthropic and OpenAI: the same model with the same prompt can produce unusable code in 20 minutes or a working game editor in 6 hours—the only variable is the harness. The course treats this as an engineering discipline, not a prompting trick, with five explicit subsystems (Instructions, State, Verification, Scope, Lifecycle) that constrain when, where, and how the agent writes code.

Key highlights

• 13 language translations of the full course material
• Reference templates for multi-turn agent pitfalls: context loss, premature task completion, scope creep
• skills/harness-creator/ scaffolds AGENTS.md, feature lists, init.sh, and verification workflows
• PDF generation via npm run pdf:build with automated release pipeline
• MIT licensed, 7,500+ stars

Caveats

• The README references OpenAI and Anthropic harness engineering posts as “core references” but does not reproduce their methodology in detail; you’ll need to read those separately
• The “6 projects” are described in the syllabus but the README does not show their scope or difficulty level
• PDF build and screenshot refresh require local Node/npm setup

Verdict Worth bookmarking if you’re running Claude Code, Codex, or similar agents on real repositories and tired of cleaning up half-finished work. Skip it if you’re looking for prompt engineering tips or model fine-tuning—this is about the filesystem, test suite, and git history around the model.