Open-source collection of battle-tested Claude Code skills for shipping safely. Includes /review-deep — a multi-agent code review that fans out four specialized reviewer agents (dependency impact, race conditions, logic & security, tenant isolation) then adversarially verifies every contestable finding with skeptic agents before reporting; /review-plan — a pre-implementation plan review that challenges assumptions, audits failure paths and ends with a Go/Adjust/Rethink verdict; /loopable — an automation scout that mines git history, scripts and Claude usage for recurring chores, ranks them, then ships the top candidate end-to-end (plan, review-plan, implement, review-deep, commit, ticket); and /init-agents — a scaffolder that detects a repo's stack and installs a tailored agent team (subagents, pipeline commands, quality hooks). One-line idempotent installer included.
Tech Stack
- Claude Code
- Multi-Agent Systems
- Prompt Engineering
- Bash
- Shell Hooks
More Projects
yad2-watcher
Self-hosted, open-source watcher for Israel's biggest classifieds site (yad2) that turns a saved car or apartment search into instant, AI-scored Telegram alerts. Your own machine fetches the search, an LLM scores every new listing 0–10 against a per-vertical rubric, and only the good deals reach you — each with a one-line reason. Ships cars + rentals adapters, a one-command launchd/cron scheduler, and a Claude Code skill that turns a plain-English brief into a real search. Python standard library only, no cloud.
Skill Factory
Open-source Claude Code tool that studies how you actually work and writes new skills for you. A read-only scanner streams your own Claude Code session transcripts and skill-usage log, isolates the multi-step tasks you still do by hand (the ones no skill has automated yet), and distills hundreds of megabytes of history into a compact digest. One Claude pass then clusters them, ranks the top 5 by how often they recur and how much manual effort they save, and drafts a complete, ready-to-install SKILL.md for each. Runs entirely on your own Claude subscription via Claude Code, reads your transcripts read-only, and never writes a skill without showing a diff and asking first. Ships with a local web UI — Generate, review 5 ranked cards, one-click install — plus a CLI and a double-clickable macOS launcher.
chaosbox
Open-source "vote-merged" playground with no maintainer in the loop: anyone opens a pull request adding a page under /contributions, the crowd votes 👍/👎 on the PR, and once it clears the vote bar with green CI and an hour of age a bot auto-merges and redeploys — shipping the contribution live in about 30 seconds. Every contribution runs sandboxed in its own iframe, so it can't touch the rest of the site, your cookies, or anyone else's page. Static HTML with zero dependencies, driven entirely by GitHub Actions.