Skill Creator (Hermes Edition)

Design and create reliable, self-enforcing skills for the Hermes Agent ecosystem.

What is this?

Skill Creator is the authoritative methodology for designing Hermes skills. It encodes the design principles, enforcement patterns, and anti-pattern catalog discovered through months of real-world skill development and debugging.

Features

• Three-layer architecture: SKILL.md → scripts/ → references/, with progressive disclosure
• Design Grill: Systematic boundary-condition analysis before writing code
• Enforcement gradients: 🤖 (mechanical) / ⚡ (load-time) / 🧠 (agent judgment) markers for rules
• skill_lint.py: Static analyzer that checks 9 quality dimensions
• 29 anti-patterns: Real failures with root causes, not theoretical best practices
• Procedural migration: Rules from prose (agent discretion) → code (mechanical enforcement)

Quick Start

# Audit an existing skill
python3 scripts/skill_lint.py <path-to-skill-directory>

# Create a new skill — follow the 5-phase operation chain:
# Phase 0: Gather Requirements → Phase 1: Design Grill → Phase 2: Static Lint
# → Phase 3: Self-Referential Test → Phase 4: Review with User

Installation

This is a Hermes skill. Copy the entire skill-creator/ directory to your Hermes skills directory:

cp -r skill-creator ~/.hermes/skills/software-development/

Or symlink it for development:

ln -s $(pwd)/skill-creator ~/.hermes/skills/software-development/skill-creator

Methodology Highlights

Attention Budget

SKILL.md < 200 lines (target), < 300 lines (hard limit). Every 100 lines beyond, Agent recall drops measurably. Domain knowledge skills get exemptions when content is operational dependency, not reference material.

Code-as-Specs

Prose instructions are wishes; pipeline gates are guarantees. If a rule can be expressed as "check X → reject/auto-fix", it belongs in scripts/, not in SKILL.md. See references/prose-vs-pipeline.md for empirical evidence.

Self-Referential Testing

Skills that define tool chains must use those tools to maintain themselves. This verifies the Agent actually follows the designed workflow, not just reads about it.

File Structure

skill-creator/
├── SKILL.md              # Core methodology (218 lines, < 300 target)
├── scripts/
│   └── skill_lint.py     # Static quality analyzer
└── references/
    ├── prose-vs-pipeline.md           # Empirical evidence for code-as-specs
    ├── hard-gate-pattern.md           # 🚨 HARD GATE design template
    ├── enforcement-gap-patterns.md    # Six patterns of rule bypass
    ├── file-checkpoint-gate.md        # Filesystem-driven step ordering
    ├── pitfall-hygiene.md             # Pitfall categorization methodology
    ├── trigger-point-enforcement.md   # Embedding prohibitions at trigger points
    ├── structured-error-pattern.md    # Error code design for data pipelines
    ├── script-as-gate.md              # Agent as analyst, not repairman
    ├── checklist-mirroring-case-study.md  # Embedded checklist enforcement
    └── absorb-external-methodology.md     # How to adopt external skill ideas

Requirements

• Python 3.8+ (for skill_lint.py)
• PyYAML (optional — improves frontmatter parsing)

License

MIT