Introduction
honey-stack is one always-on toolkit for Claude Code and Codex. It keeps a small set of rules in every session, loads skills only when a task needs them, and learns from how your sessions actually go.
Updated
What it is
It distills the best of Superpowers, Ponytail, Matt Pocock's skills, pstack and gstack into one layer, then adds RSI — recursive self-improvement. RSI learns from your corrections and can apply validated improvements on its own once you allow it.
It ships as the plugin honey in the marketplace honey-stack, with the id honey@honey-stack on both agents, plus a honey command-line tool.
Why one layer
Stacking toolkits means every session starts with roughly 25 KB of competing “YOU MUST” text. The prompts overlap, contradict each other and crowd out your code. honey-stack replaces that with about 2 KB of rules that agree with each other.
- Small. About 2 KB of always-on rules instead of about 25 KB of stacked prompts.
- Opinionated. It sizes each task, makes the smallest complete change and proves it before saying “done”.
- You choose. Learned rules apply on their own; autonomous RSI is opt-in and applies validated improvements tier by tier.
- Local and reversible. Data lives in
~/.honey/, andhoney uninstallrestores everything.
How the pieces fit
Lifecycle hooks call one entry point, bin/honey. At session start it injects the always-on rules and your learned brief. During the session a guard, a secret scan and a receipt gate run around tool calls. When the session ends, RSI reflects on what happened. The agent picks skills as tasks arrive.