An agent skill for setting up and improving harness engineering — the infrastructure that makes AI agents work effectively on your codebase.
Harness = the operating system for AI agents. Model is CPU, context window is RAM, harness is OS.
# English version
npx skills add 10xChengTu/harness-engineering/skills/harness-engineering
# 中文版
npx skills add 10xChengTu/harness-engineering/skills/harness-engineering-zhThis skill teaches your AI agent how to build and maintain the harness layer for any project — the AGENTS.md, docs/, lint rules, constraints, and evaluation systems that determine whether agents produce good or bad output.
Core principle: Start simple, add complexity only when needed. Every harness component encodes an assumption about what the model can't do alone.
| You say... | The skill does... |
|---|---|
| "Set up this project for AI agents" | Full project harness setup |
| "Create an AGENTS.md" | Scaffolds entry point + docs structure |
| "The agent keeps ignoring conventions" | Diagnoses harness gaps, not model problems |
| "Why does it keep doing X wrong?" | Identifies root cause in harness layer |
| "Make agents work better on this codebase" | Assesses & incrementally improves harness |
The skill includes 7 reference modules that the agent consults as needed:
| Module | What It Covers |
|---|---|
| Project Setup | AGENTS.md structure, docs/ directory, design notes, init scripts |
| Context Engineering | What agents see, progressive disclosure, working state management |
| Constraints & Guardrails | Linters, type systems, architecture enforcement, safe autonomy |
| Multi-Agent Architecture | Agent separation, coordination protocols, delegation patterns |
| Eval & Feedback | Testing agent output, grading, observability, feedback loops |
| Long-Running Tasks | Progress tracking, context resets, handoff artifacts |
| Diagnosis | When agents underperform — symptom → root cause mapping |
This project provides two installable skills with identical content in different languages:
# English
npx skills add 10xChengTu/harness-engineering/skills/harness-engineering
# 中文
npx skills add 10xChengTu/harness-engineering/skills/harness-engineering-zhPoor agent output is almost always a harness problem, not a model problem. When your agent ignores conventions, makes wrong assumptions, or produces inconsistent results — the fix is better context, constraints, and feedback loops, not a bigger model.
This skill encodes the patterns and anti-patterns learned from real-world agent deployments so you don't have to rediscover them.
Works with any agent that supports the Agent Skills specification, including Claude Code, OpenCode, Cursor, Codex, Cline, GitHub Copilot, and 40+ more.