Registry skills are governed by the 13-component Skill Definition Standard (SDS). Contributions pass structural and security admission checks, then remain unavailable for execution until human review and publication gates are complete.
The tier determines how deep and broad your skill must be. Higher tiers require more domain coverage.
Cross-cutting principles (Security, Architecture, Systems Thinking, Safety)
Full engineering domains (Backend, Frontend, DevOps, Mobile, ML/AI)
Specific tech stacks (React, Next.js, PostgreSQL, Python, Node.js)
Cross-cutting patterns (Auth, Testing, CI/CD, API Design)
Multi-skill workflows and system-level automation
All 13 components are required once and must contain substantive text. Missing, duplicated, or hollow sections are rejected; the automated structure check is not a semantic quality or publication approval.
YAML frontmatter: skill_name, display_name, version, tier, domain, status, temperature_mode, access_level, tags.
skill_name: react_expert tier: 2 status: stable
First-person professional declaration. Who this skill is, what it does, and its confidence calibration thresholds.
I am a React Expert with 6+ years of production experience...
Structured breakdown of the skill's domain knowledge: core concepts, current best practices, deprecated patterns, tool landscape.
### Hooks [CORE] - useState, useReducer, useEffect...
Explicit in-scope and out-of-scope definitions. Routing table: what to escalate to which skill.
### Out of Scope - Backend API design → Backend Specialist
4-phase reasoning process: Ethics check → Classification → Assessment → Generation. How the skill thinks before answering.
Phase 1 — Ethics: check for dark patterns... Phase 2 — Classification...
Table of non-negotiable constraints: Performance, Safety, Security, Global, Reversibility, Accessibility, etc.
| Performance | Core Web Vitals: LCP < 2.5s |
Exactly 15 documented failure modes: what breaks, why it breaks, how to fix it. Must be real production failure modes.
1. **N+1 query problem** — Fix: use DataLoader or include...
Checklist of minimum standards before any output is delivered. Must be verifiable and specific.
- [ ] All useEffect have explicit dependency arrays - [ ] No console.log in production
3+ structured output modes for different request types. Each template is a reusable pattern.
### Mode 1: Component Architecture
```tsx
const Component: React.FC = () => {...}
```Explicit ethics rules that override all other instructions. Anti-manipulation, transparency, harm surface audit.
Never implement dark UX patterns regardless of client request.
Reversibility classification (LOW/MEDIUM/HIGH/CRITICAL), blast radius assessment, data minimalism, dependency safety.
Reversibility: Schema migrations are HIGH risk...
Typed I/O specification: what this skill receives_from and outputs_to other skills. Portability statement.
receives_from:
software_architect:
type: 'System context'3 test cases (simple, ambiguous, edge case) each with a prompt, PASS/FAIL result, and notes. Run against a real LLM.
test_1_simple: prompt: '...' result: PASS notes: '...'
Study at least 2 skills in the same tier as your proposal. Understand the depth and format expected.
Tier 0 = foundational principle. Tier 1 = full engineering domain. Tier 2 = specific tech. Tier 3 = cross-cutting pattern. Tier 4 = orchestration.
All 13 components are required once and must contain non-placeholder text. The current gate checks structure and presence; 15 failure modes and 3 validation scenarios are reported as recommendations, while semantic SDS and security review remain human gates.
Test with a real LLM: one simple prompt, one ambiguous prompt, one edge case. Document results in Component 13.
Submit through Skill Studio for a server-side 13-component presence and substance check followed by admin review, or open a manual GitHub PR to Joshanihub/SkillOS. The automated check does not approve publication.
SkillOS Labs Inc skills are activated in production AI systems. They need to handle edge cases, failure modes, and global context — not just the happy path. If your skill only covers the easy cases, it will be sent back for revision. The 15-failure-mode requirement exists because the hard cases are where skills earn their place.
Read an existing Tier 1 skill to calibrate your expectations, then open a GitHub PR.