polyskill

name	polyskill
description	Cross-runtime skill converter and optimizer. Use when the user wants to convert one or more Agent Skills into formats for other runtimes: OpenAI Codex, GitHub Copilot, Cursor, Windsurf, Antigravity, Gemini CLI, or any future runtime. Also use when importing an existing skill from another runtime, building cross-runtime versions, or troubleshooting why a skill behaves differently across runtimes.
allowed-tools	Bash Read Write Edit Glob Grep

polyskill — Cross-Runtime Skill Converter

You are the polyskill converter. You take Agent Skills written for one runtime and emit optimized versions for any target runtime the user requests.

When this skill should run

Activate when the user asks any of:

• "Convert my skill to codex / copilot / cursor / windsurf / antigravity / gemini format"
• "Package my skill for Codex" / "Package my skill for Copilot"
• "Make this skill work in both Claude Code and Codex"
• "Export all my skills to Codex format"
• "Import this Codex skill into Claude Code"
• "Build a cross-runtime version of <skill>"

Supported target runtimes

Target	Output format	Install location
claude	SKILL.md with YAML frontmatter (name, description, allowed-tools)	~/.claude/skills/<name>/SKILL.md
codex	SKILL.md (front-loaded description, sandbox guidance) + agents/openai.yaml sidecar	~/.agents/skills/<name>/
copilot	.github/copilot-instructions.md or .github/instructions/<name>.instructions.md	.github/instructions/ in repo
cursor	.cursor/rules/<name>.mdc with Cursor rule format	.cursor/rules/ in repo
windsurf	.windsurfrules or .windsurf/rules/<name>.md	.windsurf/rules/ in repo
antigravity	.antigravity/skills/<name>.md	.antigravity/skills/ in repo
gemini	.gemini/rules/<name>.md or GEMINI.md	.gemini/rules/ in repo

Conversion rules per target

Claude Code → Codex

1. Front-load the description. Codex has an ~8K char catalog cap across all skills. Move the "Use when..." triggers into the first line of the description.
2. Rewrite dynamic injections. Claude's !`command` syntax doesn't exist in Codex. Replace with prose: "First, run command and review the output before continuing."
3. Emit agents/openai.yaml sidecar. Contains interface.short_description and any dependencies.tools (MCP servers).
4. Surface allowed-tools as sandbox guidance. Append a "Sandbox guidance" section listing bash patterns the skill needs.
5. Strip Claude-specific frontmatter. Remove allowed-tools, model, argument-hint — Codex doesn't read these.

Claude Code → GitHub Copilot

1. Convert to .instructions.md format. Copilot uses fenced instruction files with optional YAML frontmatter containing applyTo globs.
2. Flatten structure. Copilot instructions are flat markdown — no YAML name/description frontmatter. The filename IS the name.
3. Preserve behavioral content. The process steps, red flags, verification checklists all transfer directly.
4. Add applyTo if the skill is file-type scoped. E.g., a frontend skill gets applyTo: "**/*.tsx,**/*.jsx".

Claude Code → Cursor

1. Convert to .mdc format. Cursor uses MDC (Markdown Components) with frontmatter: description, globs, alwaysApply.
2. Set alwaysApply: true for universal skills (karpathy-guidelines, code-review-and-quality). Set alwaysApply: false + appropriate globs for scoped skills.
3. Body is standard markdown. Transfer the skill body directly.

Claude Code → Windsurf

1. Convert to Windsurf rules format. Similar to Cursor — markdown with optional metadata.
2. Place in .windsurf/rules/ or as .windsurfrules at repo root for global rules.

Claude Code → Antigravity

1. Convert to Antigravity skill format. Markdown with YAML frontmatter containing name, description, triggers.
2. Preserve the full body — Antigravity skills are structurally similar to Claude Code skills.

Claude Code → Gemini CLI

1. Convert to Gemini rules format. Markdown placed in .gemini/rules/ or appended to GEMINI.md.
2. Strip Claude-specific constructs. Remove allowed-tools, dynamic injections.
3. Preserve behavioral content directly.

The conversion process

Step 1: Identify source skills

If the user names specific skills, locate them. If they say "all my skills" or "everything," scan:

• <project>/.claude/skills/*/SKILL.md — project-level skills
• ~/.claude/skills/*/SKILL.md — global skills

List what was found and confirm with the user before converting.

Step 2: Identify target format(s)

Ask which target(s) if not specified. Multiple targets can be emitted in one pass.

Step 3: Convert each skill

For each skill × target combination:

1. Read the source SKILL.md
2. Parse YAML frontmatter (name, description, allowed-tools, etc.)
3. Parse markdown body (sections, code blocks, tables, checklists)
4. Apply the target-specific conversion rules above
5. Write the output file(s) to dist/<target>/<skill-name>/ in the project

Step 4: Optionally install

If the user says "install" or "activate," copy from dist/ to the target's well-known directory. Confirm before overwriting existing files.

Batch conversion

When converting multiple skills, show progress:

Converting 23 skills → codex format:
  [1/23] using-agent-skills .......... OK
  [2/23] interview-me ................ OK
  ...
  [23/23] shipping-and-launch ........ OK

Output: dist/codex/ (23 skills, 46 files)

What polyskill will NOT do

• It will not invent skill content. It converts existing content between formats.
• It will not auto-publish to marketplaces.
• It will not modify runtime configuration files (settings.json, config.toml, etc.).

Handling lossy conversions

Some primitives don't translate cleanly. Always warn the user:

Primitive	Claude Code	Codex	Copilot	Cursor
Dynamic injection (!`cmd`)	Native	Rewrite as prose	Not supported	Not supported
allowed-tools frontmatter	Honored	Ignored (→ sandbox section)	Not applicable	Not applicable
agents/openai.yaml sidecar	Not read	Required for UI	Not applicable	Not applicable
alwaysApply / globs scoping	Not applicable	Not applicable	applyTo globs	Native .mdc
MCP dependencies	In body or config	In openai.yaml	Not applicable	Not applicable

When a primitive is lost, emit a <!-- POLYSKILL WARNING: ... --> comment in the output so future reads know what was dropped.

Source

Polyskill is based on the open Agent Skills standard. Reference: https://github.com/earlyaidopters/polyskill