agent-skill-creator

name	agent-skill-creator
description	Convert any workflow, process, or tacit knowledge into a validated, reusable AI agent skill (SKILL.md + optional scripts/references/assets). Use when the user wants to: (1) turn a recurring workflow into a skill, (2) create a skill from documentation, code, transcripts, or plain English description, (3) generate a SKILL.md with proper frontmatter, structure, and validation, (4) install or share a skill across 14+ AI tools (Claude Code, Codex, Gemini, Cursor, etc.), or (5) validate or security-scan an existing skill. Not for one-off tasks that will never repeat.
metadata	{"short-description":"Turn workflows into reusable agent skills"}

Agent Skill Creator

Turn any workflow into a validated, installable AI agent skill.

What This Skill Does

Takes messy inputs (docs, links, code, PDFs, transcripts, or plain English) and generates a structured skill directory with:

• SKILL.md — proper YAML frontmatter + markdown instructions
• scripts/ — executable helpers (optional)
• references/ — documentation loaded on demand (optional)
• assets/ — templates, configs, images (optional)

Every generated skill passes validation and security scanning before delivery.

Three-Phase Pipeline

UNDERSTAND → Read all material → uncover real intent → generate internal spec
BUILD      → Structure directory → write code and docs → craft activation keywords
VERIFY     → Spec validation → security scan → block delivery if either fails

Quick Start

1. Describe your workflow

Every week I pull sales data from our CRM, clean duplicate entries, calculate
regional totals, and generate a PDF report.

Valid inputs: plain English, documentation links, existing code, API docs, PDFs, database schemas, transcripts. Combine multiple sources in one message.

2. Generated output

sales-report-skill/
├── SKILL.md          # Skill definition (activates with /keyword)
├── scripts/          # Functional Python code
├── references/       # Detailed documentation
├── assets/           # Templates, configs
├── install.sh        # Cross-platform installer (14 tools, format adapters, --all flag)
└── README.md         # Installation instructions

3. Install everywhere

# One-liner: auto-detects all tools, symlinks everywhere
./install.sh --all

# Or manual: clone to universal path
git clone https://github.com/your-org/sales-report-skill.git ~/.agents/skills/sales-report-skill

Key paths:

• ~/.claude/skills/ — Claude Code + VS Code Copilot (1.108+)
• ~/.agents/skills/ — Codex CLI, Gemini CLI, Kiro, Antigravity, and others
• ~/.codex/skills/ — OpenAI Codex
• ~/.cursor/skills/ — Cursor IDE

Run git pull once to update everywhere.

Platform Support (14 Tools)

Tier	Platforms	Behavior
Tier 1 — Native	Claude Code, Copilot, Codex CLI, Gemini CLI, Kiro, Antigravity, Goose, OpenCode	Reads SKILL.md directly
Tier 2 — Auto-adapted	Cursor (.mdc), Windsurf (.md rules), Cline, Roo Code, Trae	Installer auto-converts format
Tier 3 — Manual	Zed, Junie, Aider	Copy skill body into tool's config

Team Sharing

Simple Sharing

Agent auto-creates repo, pushes skill, generates one-liner:

git clone https://github.com/your-org/sales-report-skill.git ~/.agents/skills/sales-report-skill

Skill Registry (organizations with many skills)

Shared git repo ({team}-skills-registry) where all members publish and browse skills. No servers, no databases — just git.

Consultant model: Teach, don't build. Install tool → create registry → teach 5-step workflow → team becomes self-sustaining.

Quality Gates

Gate	Checks
Spec Validation	SKILL.md structure, frontmatter, naming, file references
Security Scan	No hardcoded API keys, no credentials, no injection patterns
Staleness Check	Review dates, dependency health, API schema drift

Run independently: scripts/validate.py, scripts/security_scan.py

Staleness Detection (3 Layers, Opt-in)

Layer	What It Does	Flag
Review tracking	Flags skills past their review date	automatic
Dependency health	HTTP-checks declared external URLs	--check-deps
Schema drift	Compares actual API response keys against expected	--check-drift

Frontmatter fields (optional — existing skills work unchanged):

reviewed: 2024-01-15
review_cycle: 90d
dependencies:
  - url: https://api.example.com/v1
    required_keys: [id, name, status]

Registry scan: stale (checks all published skills at once)

Skill Structure Standards

SKILL.md frontmatter (required)

---
name: skill-name
description: >
  What this skill does and when to use it. Be specific about triggers.
metadata:
  short-description: One-line summary
---

Body guidelines

• Keep under 500 lines; split into references/ when larger
• Use imperative/infinitive form
• Include concrete examples, not just abstract instructions
• Reference bundled files clearly: "See references/schema.md"

Security Rules

• Never hardcode API keys, tokens, or credentials in skills
• Use environment variables or 1Password CLI (op read) for secrets
• Validate all user inputs before executing scripts
• Flag injection risks in generated code
• Block delivery if security scan fails

Verification

After creating a skill, verify it works:

1. Structure check: scripts/validate.py <skill-folder>
2. Security scan: scripts/security_scan.py <skill-folder>
3. Forward-test: Use the skill on a realistic task in a fresh session
4. Install test: Run ./install.sh --all and verify it appears in all tool paths

Source

Based on https://github.com/FrancyJGLisboa/agent-skill-creator (v5.0.0, MIT).