OpenMythos Skill

OpenMythos-Skill: A Claude Skill for Kye Gomez's OpenMythos Repo

A [Claude skill](https://docs.claude.com) that gives Claude careful, repo-specific knowledge of **OpenMythos** — Kye Gomez's open-source PyTorch reconstruction of a hypothesized Recurrent-Depth Transformer architecture. When this skill is loaded, Claude acts like a senior engineer who has read every line of `open_mythos/main.py`: it knows the architectural invariants, the conventions, the debugging playbook, and the papers the repo cites.

This is a **third-party harness for working with OpenMythos**. It does not modify the OpenMythos codebase, does not redistribute its source, and is not affiliated with either the OpenMythos project or Anthropic.

---

Credit where it's due

All the substance this skill knows comes from **Kye Gomez**'s work on OpenMythos:

• Repo: https://github.com/kyegomez/OpenMythos
• License: MIT (Copyright © 2026 Kye Gomez)
• Twitter/X: @kyegomezb

If you're going to use this skill, go star and read [the upstream repo](https://github.com/kyegomez/OpenMythos) first. The README there is genuinely one of the better explainers of looped-transformer theory I've read — LTI stability via `A = exp(-exp(...))`, loop-index embedding, ACT halting, Parcae scaling laws, the memorization-reasoning tradeoff, all laid out clearly with citations. This skill is a harness around that content; the content itself is Kye's.

---

What the skill actually does

When loaded, the skill triggers whenever Claude sees signals that the user is working on OpenMythos: filenames like `open_mythos/main.py`, imports like `from open_mythos.main import OpenMythos, MythosConfig`, symbol names (`MythosConfig`, `RecurrentBlock`, `LTIInjection`, `ACTHalting`, `MoEFFN`, `MLAttention`, `GQAttention`), variant helpers (`mythos_1b` through `mythos_1t`), and debugging symptoms specific to looped training (step-reproducible loss spikes, residual explosion, overthinking dri