Developer Bennyoooo has released Skillmaxxing on Hacker News, an open-source plugin that transforms any coding agent into a self-improving system by automatically capturing and refining reusable workflows as skills. The project installs with a single npm command—npm i -g skillmaxxing && skillmaxxing plugin install—and requires only Node.js 20 or higher. Once installed, the plugin works silently in the background across Claude Code, Codex, and compatible agents without any manual invocation.
How the Self-Evolution Loop Works
Skillmaxxing installs two lightweight hooks into your agent's lifecycle. The SessionStart hook injects standing guidance that nudges the agent to crystallize reusable work when it appears. When a session ends via the Stop hook, the system counts tool calls in the transcript; if enough new work has accumulated past a configurable threshold (defaulting to 12), it forks a background reflector process that reviews the session and—only when warranted—creates one new skill or improves an existing one. This mirrors the architecture of the Hermes Agent's self-improvement loop, adapted for agents like Claude Code and Codex.
Two Modes for Different Workflows
The auto mode (default) runs reflection entirely in the background using a headless agent process (claude -p, restricted to skill tools), letting you continue working uninterrupted while your agent gets smarter. The nudge mode instead reminds the agent to crystallize workflows itself during the active session—ideal for developers who want zero extra processes or full visibility into what skills get created. You can switch modes with a flag: skillmaxxing plugin install --mode nudge --threshold 12. Either way, every skill written by the reflector is marked trusted: false and never auto-executes without your explicit approval.
Trust, Safety, and Reversibility
Skillmaxxing takes a conservative approach to untrusted code. New or improved skills are always created with trusted: false, meaning the sandbox refuses to run their contents unless you grant permission explicitly via --allow-exec. The optimization loop is eval-gated—candidates only get promoted on strict score wins with no regression, all prior versions are retained indefinitely, and any change stays reversible. The background reflector is also recursion-guarded and fully detached from your session, so it can never trigger itself or interfere with active work.
CLI Primitives Beyond the Plugin
Everything in the plugin layer builds on two core CLI commands that you can call directly. skillmaxxing skillify --draft draft.json stages a workflow, runs smoke tests, presents results for review, and commits only after approval—useful for manually crystallizing complex processes. The skillmaxxing optimize command provides an eval-gated improvement loop with bounded edits and full validation, making it safe to iteratively improve existing skills over time.
Roadmap Includes Discovery and Team Sharing
The v1 release focuses on hook-driven auto-creation and eval-gated improvement across Claude Code and Codex. CLI commands for discovering public skills (skillmaxxing discover) and team workspaces (skillmaxxing workspace) already exist but are intentionally held out of the plugin surface to keep installation friction-free. Those features will land in a future plugin update, along with collaborative optimization capabilities.
The Bottom Line
Skillmaxxing solves a real pain point: every session starts fresh because agents forget everything. If you run Claude Code or Codex daily and find yourself re-explaining the same repo quirks or re-deriving the same migration patterns, this one-line install is worth five minutes of your time—though you'll want to audit auto-created skills before blindly trusting them in production pipelines.