Agent Skills
4 hours ago
- #AI coding agents
- #agent harness
- #software engineering practices
- Agent Skills is a framework to enforce senior engineer practices (like specs, tests, reviews) in AI coding agents, which typically skip these steps by default.
- Skills are markdown workflows with frontmatter, injected contextually—focusing on actionable steps with exit criteria, not reference documentation.
- The project organizes twenty skills around six SDLC phases: Define, Plan, Build, Verify, Review, Ship, plus a /code-simplify command.
- Five key design principles: process over prose, anti-rationalization tables, non-negotiable verification, progressive disclosure, and scope discipline.
- Skills incorporate Google engineering practices (e.g., Hyrum's Law, test pyramid, 100-line PRs) to ensure reliability at scale.
- Usage modes include installing via marketplace, dropping markdown into tools, or reading skills as specifications for team practices.
- Anti-rationalization tables preempt excuses for skipping workflows, a pattern applicable to human teams to prevent engineering decay.
- Verification requires concrete evidence (e.g., tests passing, reviewer sign-off)—'seems right' is insufficient for task completion.
- Progressive disclosure loads only relevant skills into context to maintain performance, avoiding token overload.
- The framework targets long-running agents, where skipped steps accumulate into major issues, and is portable across AI coding tools.