Evidence before confidence
Separate documented claims from executable, enforced, and maintained evidence. Keep unknowns and command limitations visible.
MetaSkills 1.2 · Build agents, then validate their harness
Build better agents. Validate the system around them.
MetaSkills packages portable practices for building agents and checking whether their instructions, environment, permissions, state, verification, observability, recovery, and orchestration are trustworthy.
The new harness-validator stays read-only, separates claims from executable evidence, surfaces blockers that averages must not hide, and recommends the smallest effective improvements.
npx skills add https://github.com/smota/metaskills --skill harness-validator --full-depth
Why it matters
Agent capability should be portable and inspectable.
Great agent workflows are hard won: guardrails, context boundaries, examples, and validation steps are often trapped in private chats or one developer's habits.
MetaSkills turns that knowledge into lightweight, readable files that live in Git. Install a skill, adapt it to your project, inspect the practice, and improve it openly without adopting a proprietary runtime.
What is new in 1.2
Validate the harness, not just the agent.
Harness Validator applies 78 canonical checks across nine categories, while adapting rigor to project risk, lifetime, architecture, and autonomy.
01Knowledge
02Environment
03Security
04Scope
05Continuity
06Verification
Blockers stay visible
Unsafe access, exposed secrets, missing completion criteria, unreproducible setup, and unsupported cross-boundary claims never disappear into an average.
Right-sized recommendations
Mark advanced controls not applicable when justified, detect contradictions and over-detail, and prefer the smallest improvement that reduces real risk.
Harness agnostic, capability aware
Works as files. Upgrades when your harness has power.
The canonical agent stays portable. Harness adapters describe how to use platform features when present, and what to do when they are missing.
Portable core
`AGENT.md`, knowledge sources, tool/action contracts, evals, guardrails, and validation live in Git and can move across projects.
Optional adapters
Subagents, async jobs, MCP/OpenAPI actions, memory, and monitoring become adapter capabilities with explicit manual fallback paths.
Skill catalog
Choose the reusable practice you need.
Each skill is self-contained and can be installed into an agent-specific skill directory such as `.agents/skills`, `.claude/skills`, `.codex/skills`, or another supported target.
agent-builder
Build a portable agent from a goal.
Use when you want a copilot, coding assistant, tool-enabled reviewer, or knowledge-focused onboarding agent. Version 1.1 adds a maturity model from Level 0 to Level 5, harness adapters for Pi, Claude, Codex, Gemini, and Microsoft Foundry, plus templates for handoffs, jobs, feedback, benchmarks, and runtime capabilities.
npx skills add https://github.com/smota/metaskills --skill agent-builder --full-depth
skill-creator
Turn a workflow into an installable skill.
Use when you want to create, evaluate, improve, benchmark, or evolve reusable AI agent skills.
npx skills add https://github.com/smota/metaskills --skill skill-creator --full-depth
harness-validator · New in 1.2
Audit the system around a coding agent.
Use when you need a read-only, evidence-first review of instructions, environment, security, task control, state continuity, verification, observability, clean handoffs, or autonomous orchestration.
npx skills add https://github.com/smota/metaskills --skill harness-validator --full-depth
Proof by example
See what you can build.
Examples show the starting prompt, selected skill, expected output structure, before/after value, assumptions, and validation checklist.
Release notes agent
Draft release notes from issues, pull requests, and changelog fragments.
Support triage skill
Package support classification, escalation, privacy, and response workflow.
Migration reviewer
Create an agent that reviews database changes for release blockers and risk.
For builders
Works with the tools you already use.
MetaSkills is vendor-neutral and file-based. It can support OpenAI/Codex-style coding agents, Claude, GitHub Copilot, Gemini, and other agent harnesses that read repository instructions.
FAQ
Built for adoption, not dependency.
Quick answers for OpenAI users, GitHub maintainers, and teams evaluating agent tooling.
What is MetaSkills?
MetaSkills is an open-source collection of installable skills for creating agents, skills, and reusable agent workflows.
Is this a hosted platform?
No. It is not a runtime, control plane, deployment service, or management system. The assets are readable files intended to live in your repository.
How do I start?
Run npx skills add https://github.com/smota/metaskills --list --full-depth, choose a skill, install it, and ask your coding agent to use it by name.
How do adopting projects update?
Re-run the install command from the adopting project root, then review the diff before accepting upstream template, checklist, reference, or adapter changes.
How can the community contribute?
Open a GitHub issue or PR with a skill, example, toolset, component reference, or docs improvement. Start with the contribution guide.