Superpowers 3.3 Ports SKILL.md Workflows to Codex

Superpowers 3.3.0 runs SKILL.md agent workflows on OpenAI Codex with a Claude-to-Codex mapping, bootstrap, and helper scripts. Installer available — warns it downloads and executes code (remote execution risk).

•
Codex
tool cover

TL;DR

  • Superpowers 3.3.0 ported to OpenAI Codex: superpowers-codex script and updated bootstrap that populates ~/.codex/AGENTS.md.
  • Claude-to-Codex dictionary added to translate tool names and behaviors (e.g., TodoWrite → update_plan) so existing SKILL.md can run without rewriting.
  • Autodiscovery of personal skills via ~/.codex/skills/, allowing overrides on name collisions.
  • Retains core pieces: SKILL.md files, a skill runner, and a bootstrap; translation layer and strict bootstrap required because Codex follows instructions literally.
  • Installer and safety warning: installer available at https://raw.githubusercontent.com/obra/superpowers/refs/heads/main/.codex/INSTALL.md; running it downloads and executes code and edits ~/.codex/AGENTS.md (prompt injection / remote code execution risk).
  • Full write-up and technical details: https://blog.fsck.com/2025/10/27/skills-for-openai-codex/

Superpowers lands on Codex: SKILL.md workflows now work outside Claude

The Superpowers workflow—an agent-centric system built around reusable SKILL.md process documents—has been ported from Anthropic’s Claude Code to OpenAI’s Codex CLI in Superpowers 3.3.0. The move adapts the same instrumented “binder” idea for agentic development to a model that lacks plugins and hooks by default, using a bootstrap, tool-mapping layer, and a small helper script to bridge the gaps.

What changed

  • The Superpowers repo and its SKILL.md system are now runnable under Codex, with a superpowers-codex script and an updated bootstrap that populates ~/.codex/AGENTS.md.
  • A Claude-to-Codex dictionary was added to translate tool names and expected behaviors (for example, mapping TodoWrite to update_plan) so existing skills can run without rewriting them.
  • Support for a ~/.codex/skills/ directory was added so personal skills can be autodiscovered and override Superpowers skills when names collide.

How the system works

Superpowers retains the same three core pieces:

  • SKILL.md files: human-readable, agent-executable descriptions of when and how to use a skill.
  • A skill runner: a tool that instructs the agent to read and follow a SKILL.md.
  • A bootstrap: announces available skills to the agent at startup.

Because Codex is highly literal in following instructions, the translation layer and the strict bootstrap rules were essential. The author reports that Codex “really, really likes to follow instructions,” and once the mapping and discovery pieces were in place, agents behaved as intended—loading brainstorming and TDD workflows from skills before starting work.

Install note and safety warning

An immediate way to try the port is to fetch the installer from the repository, for example: Fetch https://raw.githubusercontent.com/obra/superpowers/refs/heads/main/.codex/INSTALL.md and follow the instructions.

The author explicitly warns that running that command causes the agent to download and execute code, editing ~/.codex/AGENTS.md—a literal prompt injection and remote code execution action. The risk and mechanics are explained in the original post.

For more detail on the bootstrap tricks, tool mappings, and behavioral differences observed between Claude and Codex, read the full write-up: https://blog.fsck.com/2025/10/27/skills-for-openai-codex/

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community

Related Articles

openai cover

GPT‑5.1‑Codex‑Max Delivers Compaction, Multicontext and Long Coding Workflows

GPT‑5.1‑Codex‑Max adds compaction and token‑efficient reasoning to Codex, enabling coherent multi‑hour project coding across CLI, IDE, and cloud. It runs sandboxed by default with monitoring and safety controls.

1 shared tag
Run Codex CLI on Mac Against gpt-oss:120b via Tailscale

Run Codex CLI on Mac Against gpt-oss:120b via Tailscale

Simon Willison shows how to run OpenAI’s Codex CLI on a Mac while offloading the actual AI model to a GPT-OSS 120B instance running on his DGX Spark via Tailscale and Ollama. It works—even for building a Space Invaders game—but it’s more of a fun hack than a daily driver.

1 shared tag
Practical Playbook for Agentic Engineering: Talk Directly to Agents

Practical Playbook for Agentic Engineering: Talk Directly to Agents

A concise, practical guide to agent-driven development for large TypeScript/React codebases. Focuses on direct CLI interaction with gpt-5-codex agents, atomic commits, blast-radius edits, and lightweight planning.

1 shared tag