Write A Skill

Write A Skill: Progressive Disclosure & Agent Skill Authoring

When to Use This Skill

Triggers — load this skill when:

• A new skill is being created and must pass structure and description gates
• An existing skill is bloated, untriggered, or inconsistently structured
• You are reviewing a skill contribution before merge

Route elsewhere when:

• Repository-wide validation, eval, compliance and mirror-sync gates -> the repository-root scripts/ directory
• Contribution conventions -> CONTRIBUTING.md

Attribution

Derived from Matt Pocock's write-a-skill (MIT-licensed).

1. Non-Negotiable Skill Design Laws

1. The Description is King: The description field in the frontmatter is the only piece of information the model reads when deciding whether to activate the skill. It MUST be written in the third person and specify explicit triggers starting with "Use when...".
2. The 100-Line Ceiling: Keep the main SKILL.md concise (< 100-150 lines). Avoid overwhelming context windows. Offload deep dive runbooks, deep schemas, and multi-page catalogs into references/ or scripts/.
3. Progressive Disclosure: Only the frontmatter is in the initial system prompt. The model reads SKILL.md on-demand, and reads files under references/ only when resolving deep edge cases.

2. The 3-Phase Skill Authoring Workflow

Phase 1: Gather

• Define the exact scope and operational boundaries of the skill.
• Collect production-tested configurations, manifests, commands, and failure modes.
• Identify the exact keywords and scenario triggers ("Use when...").

Phase 2: Draft

• Scaffold directory structure:

  skills/<category>/<skill-name>/
  ├── SKILL.md                # Entry point with YAML frontmatter (<100-150 lines)
  ├── scripts/                # Standalone helper CLI scripts or validators
  └── references/             # Deep references, architecture notes, cheat sheets
  

• Write concise, actionable sections:

  1. Quick Reference / Decision Matrix
  2. Production Configurations / Commands
  3. Anti-Patterns & Safety Checks

Phase 3: Review Checklist (Matt Pocock's 6 Gates)

• Description Gate: ≤ 1024 chars, third person, contains explicit "Use when..." triggers.
• Length Gate: Main SKILL.md is tight, actionable, and delegates deep references.
• Executable Code Gate: Code snippets are copy-pasteable without placeholders like TODO or <fill_in>.
• Anti-Pattern Gate: Explicitly lists "Don'ts" and failure modes.
• Portability Gate: Interoperable across Antigravity, Claude Code, Cursor, and Codex.
• Validation Gate: Passes validate-skills.py and skill_review_checklist_runner.py.

3. Bundled Validation Tools

Run the validation suite locally:

# Run review checklist runner on a skill
python3 skills/productivity/write-a-skill/scripts/skill_review_checklist_runner.py skills/devops-core/mid-level-automation/terraform-iac-modules/SKILL.md

# Validate description trigger compliance
python3 skills/productivity/write-a-skill/scripts/skill_description_validator.py skills/devops-core/mid-level-automation/terraform-iac-modules/SKILL.md

# Validate folder structure
python3 skills/productivity/write-a-skill/scripts/skill_structure_validator.py skills/devops-core/mid-level-automation/terraform-iac-modules/