Self-Improving Skills
UniM0cha/self-improving-skills
Hermes Agent-style skill self-improvement plugin for Claude Code, Claude Cowork, Codex and ChatGPT work
Self-Improving Skills is a plugin that brings the Hermes Agent self-improvement loop to Claude Code and similar tools. It turns hard-won workflow lessons into reusable SKILL.md files, validates skill edits and archives stale skills.
What Self-Improving Skills does
The repository ports the idea behind Hermes Agent's procedural memory and curator loop: finish a complex task, distill what worked, save or patch a skill, and find it again next session. It detects complex work from transcript and tool-call signals, then writes or patches a SKILL.md under ~/.claude/skills through a dedicated subagent. Since v0.13.0 the default is background distillation: a Stop hook queues a job in SQLite and a detached worker runs a headless claude -p session, with an in-turn nudge as the fallback.
Safety is layered. The background session runs with reduced tools and no Bash, permission deny rules and a per-job budget cap, and a skill guard snapshots the skills directory, validates every touched SKILL.md and rolls back anything malformed or out of scope. A curator marks unused agent-created skills stale after 30 days and archives them after 45, and /curate-skills runs an LLM curation pass. The repository holds four variants, for Claude Code, Claude Cowork, Codex CLI and ChatGPT desktop Work.
Key features
- Detects complex work from transcript and tool-call signals
- Background distillation through a queued job and a detached claude -p worker
- Skill guard with backups, validation and automatic rollback
- Usage telemetry kept in skill_usage.json
- Curator that archives unused generated skills after 45 days
- Four variants: Claude Code, Claude Cowork, Codex CLI and ChatGPT Work
When to use it
- Capturing a repeatable debugging technique as a skill after a hard session
- Keeping an auto-generated skills folder from filling with stale entries
- Bringing Hermes-style learning to a Claude Code workflow
Who it is for: Claude Code, Codex and ChatGPT users who like the Hermes Agent learning loop and want similar behavior in those tools.
How it fits with Hermes Agent
Not a Hermes plugin: it ports the Hermes Agent self-improvement idea to other tools and names Nous Research's Hermes Agent as its inspiration.
Note: It targets Claude Code, Codex and ChatGPT rather than running inside Hermes Agent.
FAQ
What is Self-Improving Skills?
Self-Improving Skills is a plugin that adds a Hermes Agent-style learning loop to Claude Code. It watches for complex work, distills the reusable technique into a SKILL.md and curates stale skills.
Does Self-Improving Skills run inside Hermes Agent?
No. It targets Claude Code, Claude Cowork, Codex CLI and ChatGPT desktop Work, and borrows the approach of Hermes Agent, which already has skills and a curator loop of its own.
Is Self-Improving Skills free and open source?
Yes. The repository is released under the MIT license.
Similar skills for Hermes Agent
All skillsTrading skill built on 23 Chinese hot-money traders' methods for A-share short-term analysis
LeoYeAI Seedance Skills28 agent skills for writing Seedance 2.0 video prompts, from single clips to multi-clip sequences
AkoliteZA Superpowers for Hermes Agent and ChatGPTUnofficial 14-skill coding workflow pack adapted from Superpowers for Hermes Agent and ChatGPT
shantanugoel askjevHermes skill that calls TypeSafe Jev as a typed tiebreaker returning Choice, Score or Noul answers
arimanyus Hermes MerchantPortable skill set that scrapes ML/AI job boards, scores roles and auto-fills Greenhouse applications
MustBeSimo Web Design StudioFree MIT skill for coding agents that builds distinctive websites with 3D, motion and fallbacks
Related guides: What is Hermes Agent? · SOUL.md for Hermes Agent: what it is and how to write one