How to INSTANTLY Run ANY Skill in Claude + Codex

How to INSTANTLY Run ANY Skill in Claude + Codex — Mark Kashef

⏱ ~11 min 🎤 Mark Kashef 🏷 Skills 🏷 PolySkill 🏷 Claude Code 🏷 Codex 🏷 Portability

Skill Portability Problem 0:00

AI coding agents like Claude Code and OpenAI's Codex CLI are increasingly powerful — but they each have their own way of defining "skills" (reusable instructions that guide the agent's behavior).

  • Claude Code uses CLAUDE.md files — markdown instructions placed in the project root
  • Codex CLI uses AGENTS.md or system prompt configuration
  • Skills written for one tool don't work in the other — you have to rewrite them
The problem: You've spent hours crafting a perfect skill for Claude Code. Now you want to try Codex CLI. Rewrite everything? Copy-paste and pray? Neither option is great.

Why Same Skill Behaves Differently 1:34

Even if you manually port a skill between agents, the behavior often differs. Mark explains why:

  • Context injection: Claude Code reads CLAUDE.md at the start of every session; Codex uses AGENTS.md differently in its pipeline
  • Tool availability: Each agent has different built-in tools — file editing, search, terminal access vary
  • Instruction parsing: Each model interprets skill instructions through its own training — same words, different execution
  • Execution model: Claude Code is interactive (back-and-forth); Codex can run autonomously in sandboxed environments

These differences mean that a "skill" isn't just text — it's text tuned for a specific agent's behavior model.

Anatomy of an Agent Skill 3:27

Mark breaks down what makes up a well-crafted agent skill:

┌─────────────────────────────┐ │ Trigger / When to Use │ ← When should the agent activate this skill? │ Steps / Instructions │ ← Ordered actions the agent should take │ Constraints / Rules │ ← Guardrails and boundaries │ Tool Preferences │ ← Which tools to use (and avoid) │ Output Format │ ← How to present results │ Agent-Specific Hints │ ← Tuning for the specific agent runtime └─────────────────────────────┘
  • The first five layers are universal — they describe what to do regardless of which agent runs them
  • The last layer is agent-specific — it tunes how the instructions are expressed for a particular runtime
  • PolySkill works by separating these layers and translating only the agent-specific parts

PolySkill: Universal Adapter 5:59

PolySkill is the tool Mark built to solve the portability problem. It acts as a universal adapter between agent skill formats:

  • Input: A skill written for any supported agent (Claude Code, Codex, Hermes, Cursor, etc.)
  • Processing: Extracts the universal skill logic from agent-specific formatting
  • Output: Regenerates the skill in the target agent's format with appropriate tuning
Key feature: PolySkill doesn't just format-convert. It adapts — rewriting tool preferences, adjusting instruction styles, and adding agent-specific hints that make the skill feel native.

Supported conversions include:

  • CLAUDE.md ↔ AGENTS.md (Claude Code ↔ Codex)
  • CLAUDE.md ↔ .cursorrules (Claude Code ↔ Cursor)
  • Any format → Hermes SKILL.md and vice versa

Live Demo: Two-Way Conversion 7:22

Mark demonstrates PolySkill converting a real skill in both directions:

  • Claude → Codex: Takes a CLAUDE.md skill for test-driven development, converts it to AGENTS.md format with Codex-appropriate instructions
  • Codex → Claude: Takes a Codex AGENTS.md skill for code review, converts it to CLAUDE.md with Claude-native instruction patterns
  • Both conversions preserve the core logic while adapting the presentation layer

🔵 Claude Code (CLAUDE.md)

  • Markdown with headers and bullet points
  • References bash, Read, Write tools
  • Interactive confirmation style
  • Placed in project root or ~/.claude/

🟢 Codex CLI (AGENTS.md)

  • Structured with sections and directives
  • References shell, file_edit, file_read
  • Autonomous execution style
  • Placed in project root or global config

When to Use Which Provider 8:03

Mark closes with practical guidance on when to use each coding agent — and why skill portability matters for making that choice freely:

  • Claude Code: Best for interactive development where you want to steer the agent and review changes in real-time
  • Codex CLI: Best for autonomous batch tasks — run it, walk away, come back to results
  • Cursor: Best when you want IDE-integrated AI assistance with visual diff review
  • Hermes: Best for orchestration — combining multiple tools and agents into workflows
The meta-point: With PolySkill, you're no longer locked into one agent. Write your skill once, convert freely, and pick the best tool for each job — your accumulated knowledge travels with you.