This is a set of language agnostic agent skills focused on improving context management and enforcing good practices while keeping you in the loop to ensure alignment.
- Semble — fast code search for AI agents. Skills use semble to locate relevant code without grepping blindly. Install via
pip install sembleand add it as an MCP server in your Claude Code config.
Run this command from the root of the repo to symlink the skills into your Claude Code skills directory (change the path if using another harness):
mkdir -p ~/.claude/skills && ln -sf "$(pwd)"/skills/*/ ~/.claude/skills/- caveman — local proxy and hooks that compress agent input and output to reduce token usage
- rtk — Rust Token Killer, a CLI proxy that cuts token usage 60–90% on common dev operations
Breaks any multi-file feature or refactor into commit-sized phases — each with a clear goal, concrete file changes, success criteria, and quality gates. Claude pauses after every phase for your review before committing, so you stay in control of large changes while ending up with a clean, reviewable git history.
Compacts the current conversation into a structured briefing for a fresh agent to pick up. Use this instead of /compact when you want a clean slate — the new session inherits only what matters, rather than a compressed version of everything.
Audits and syncs documentation against the actual code. Works across READMEs, docs/, inline docstrings, and agent context files (AGENTS.md, CLAUDE.md). Detects drift via broken references and git age gaps, writes all findings before touching any files, then fixes stale content, removes cruft, and fills gaps.
Creates or updates an AGENTS.md hierarchy for a codebase — auto-detects whether to bootstrap from scratch or refresh existing docs. Mines the repo for gotchas, conventions, and build/test commands that would otherwise trip up an agent working cold.
Writes a thorough test suite for a given piece of code. Discovers the project's existing test framework and conventions, then builds a matrix covering normal behaviour, boundary conditions, error cases, and state transitions — verifying through public interfaces rather than implementation details.