Skill Specification

Formal specification for evolutionarily robust agent skills.

Why This Matters

Skills that follow the spec work across:

Claude Code (Anthropic)
Codex CLI (OpenAI)
Cursor (Anysphere)
Amp (Sourcegraph)
Letta (memGPT)
Future agents

Non-compliant skills break silently or fail validation.

SKILL.md Schema

---
name: skill-name              # REQUIRED: lowercase, hyphens, 1-64 chars
description: What and when    # REQUIRED: max 1024 chars, no < or >
license: Apache-2.0           # optional
compatibility: Requires git   # optional, max 500 chars
metadata:                     # optional: custom key-value pairs
  trit: 0
  author: bmorphism
  version: "1.0"
allowed-tools: Bash Read      # optional, experimental
---

# Body content (Markdown)

Field Constraints

| Field | Required | Rules | |-------|----------|-------| | name | ✓ | [a-z0-9-]+, no --, no leading/trailing -, max 64 | | description | ✓ | 1-1024 chars, no < or >, includes WHEN to use | | license | ✗ | Short name or file reference | | compatibility | ✗ | Environment requirements, max 500 | | metadata | ✗ | Arbitrary k:v for custom fields | | allowed-tools | ✗ | Space-delimited tool names |

Evolutionary Robustness Patterns

1. Progressive Disclosure

Level 1: name + description (~100 tokens) - loaded at startup
Level 2: SKILL.md body (<5000 tokens) - loaded on activation
Level 3: scripts/, references/, assets/ - loaded on demand

Keep SKILL.md under 500 lines. Move details to references/.

2. Cross-Platform Compatibility

# BAD - platform-specific
allowed-tools: claude_desktop_mcp

# GOOD - generic capability
compatibility: Requires MCP server access

3. Self-Validation Hook

Include validation in your skill:

# scripts/validate.sh
skills-ref validate "$(dirname "$0")/.."

4. Semantic Versioning in Metadata

metadata:
  version: "2.1.0"
  breaking-changes: "v2.0 changed API"

5. Triadic Classification (GF(3) Extension)

For plurigrid/asi skills:

metadata:
  trit: -1   # MINUS: verification, constraint
  trit: 0    # ERGODIC: balance, mediation
  trit: +1   # PLUS: generation, exploration

Conservation: Σ trits ≡ 0 (mod 3) across compositions.

Directory Structure

skill-name/
├── SKILL.md           # Required
├── scripts/           # Executable code
│   └── main.py
├── references/        # Additional docs
│   └── REFERENCE.md
└── assets/            # Static resources
    └── template.json

Validation Commands

# Official validator
skills-ref validate ./my-skill

# Codex-rs validator
python3 codex-rs/core/src/skills/assets/samples/skill-creator/scripts/quick_validate.py ./my-skill

# Batch validate
for d in skills/*/; do skills-ref validate "$d"; done

Common Failures

| Error | Fix | |-------|-----| | No YAML frontmatter | Add --- delimiters | | Unexpected keys | Move to metadata: | | Angle brackets in description | Remove < and > | | Name not hyphen-case | Lowercase, hyphens only | | Description too long | Max 1024 chars | | YAML colon in value | Quote the string |

Evolution Strategy

Start minimal - name + description + one paragraph
Add scripts/ when automation helps
Add references/ when body exceeds 300 lines
Add metadata for custom classification
Validate on every commit via CI

References

Scientific Skill Interleaving

This skill connects to the K-Dense-AI/claude-scientific-skills ecosystem:

Graph Theory

networkx [○] via bicomodule
- Universal graph hub

Bibliography References

category-theory: 139 citations in bib.duckdb

Cat# Integration

This skill maps to Cat# = Comod(P) as a bicomodule in the equipment structure:

Trit: 0 (ERGODIC)
Home: Prof
Poly Op: ⊗
Kan Role: Adj
Color: #26D826

GF(3) Naturality

The skill participates in triads satisfying:

(-1) + (0) + (+1) ≡ 0 (mod 3)

This ensures compositional coherence in the Cat# equipment structure.