archive228/loopkit

Loopkit is a drop-in set of files that gives a coding agent, such as Claude Code, Cursor, Codex, or Gemini CLI, a structured way of working: plan, act, then verify. It installs into a project's .claude folder with a single install script and does not run as a background service or daemon. Everything it adds is plain files on disk: settings, hooks, a verifier subagent, a loop runner script, and a library of 49 small skills. Each skill is a short markdown document with a YAML header describing what it does and when it should load. Skills only activate when a task actually triggers them, so the agent is not carrying every possible instruction at once. The skills are grouped into ten tracks covering things like agent behavior, long running multi session projects, debugging, and security review. The core loop the project enforces is plan, act, verify, and then revise if verification fails. Planning skills such as spec first and context budget load before work starts. Action skills for tasks like debugging, testing, refactoring, or git operations load while work happens. A separate verifier subagent, described as adversarial, checks the resulting change afterward, assuming it is broken until proven otherwise, and returns its findings as JSON rather than proposing fixes itself. A project using loopkit ends up with a small standing context file limited to 60 lines, a permissions file that allows safe read only commands while blocking risky ones like force pushes, a hooks folder that reminds the agent of the loop shape when a session starts, an MCP configuration file wired to GitHub and Context7 by default, and a memory file meant to be pruned regularly so it does not accumulate stale notes. The project describes itself as opinionated but not a methodology to learn or a vendor lock in. Since it is plain markdown and shell files under an MIT license, a user can keep only the pieces they want and discard the rest. It is aimed at developers who already use an AI coding agent and want a lightweight, file based structure for keeping that agent's work reviewable and verified rather than free form.