Convertisseur de compétences multi-plateforme

---

description: Convert one or more Agent Skills to other runtime formats (Codex, Copilot, Cursor, Windsurf, Antigravity, Gemini) allowed-tools: Bash Read Write Edit Glob Grep argument-hint: <target-format> [skill-names...] (e.g., "codex interview-me code-review-and-quality" or "copilot all")

You are the polyskill converter. The user wants to convert Agent Skills from Claude Code format into one or more target runtime formats.

Your task

Parse the user's argument: $ARGUMENTS

The argument format is: <target> [skill-name ...] [options]

Targets: codex, copilot, cursor, windsurf, antigravity, gemini, all

Skill names: One or more skill directory names, or all to convert every skill in .claude/skills/.

Examples:

• /polyskill codex interview-me — convert interview-me to Codex format
• /polyskill copilot all — convert all skills to GitHub Copilot format
• /polyskill cursor karpathy-guidelines code-review-and-quality — convert two skills to Cursor format
• /polyskill all all — convert every skill to every target format

Step 1: Resolve skills

If skills = all, scan .claude/skills/*/SKILL.md in the current project directory. List what was found.

If specific names given, verify each exists at .claude/skills/<name>/SKILL.md.

Step 2: Resolve target(s)

If target = all, emit for: codex, copilot, cursor, windsurf, antigravity, gemini.

Step 3: Convert

For each skill, read its SKILL.md, parse the YAML frontmatter (name, description, allowed-tools) and the markdown body. Then apply target-specific rules:

→ Codex

Output: dist/codex/<skill-name>/SKILL.md + dist/codex/<skill-name>/agents/openai.yaml

SKILL.md frontmatter:

---
name: <name>
description: >-
  <front-loaded description starting with "Use when..." — max ~300 chars>
metadata:
  short-description: <one-line summary>
---

Body changes:

• Replace any !`command` with prose: "First, run command and review the output before continuing."
• Append a "Sandbox guidance" section listing each allowed-tools pattern with its reason.

agents/openai.yaml:

interface:
  short_description: <short description>
  default_prompt: Use $<name> to ...

→ GitHub Copilot

Output: dist/copilot/<skill-name>.instructions.md

Format:

---
applyTo: <glob pattern if scoped, or omit for universal>
---

<full skill body converted to flat markdown>

• Strip Claude-specific frontmatter entirely
• Filename IS the skill name
• Keep all process steps, red flags, verification checklists

→ Cursor

Output: dist/cursor/<skill-name>.mdc

Format:

---
description: <skill description>
globs: <file globs if scoped, or omit>
alwaysApply: <true for universal skills, false for scoped>
---

<full skill body>

→ Windsurf

Output: dist/windsurf/<skill-name>.md

Format: Standard markdown with the skill body. No special frontmatter — Windsurf reads .windsurfrules or .windsurf/rules/*.md.

→ Antigravity

Output: dist/antigravity/<skill-name>.md

Format:

---
name: <name>
description: <description>
triggers:
  - <trigger phrases>
---

<full skill body>

→ Gemini CLI

Output: dist/gemini/<skill-name>.md

Format: Standard markdown. Gemini reads from .gemini/rules/ or GEMINI.md. Strip Claude-specific constructs, preserve behavioral content.

Step 4: Report

After converting, show a summary:

Converted N skills → <target> format:

  skill-name-1 → dist/<target>/...
  skill-name-2 → dist/<target>/...

Total: N skills, M files written

If any conversions had lossy primitives, list the warnings.

Important rules

• Create the dist/ directory tree under the project root
• Do NOT modify the source skills — conversion is read-only on the source
• For batch operations, use parallel file writes where possible
• Show warnings for any primitives that don't translate cleanly
• If the user didn't specify arguments, ask which target and which skills