inwo. Headroom sits between your AI agent and the model, compressing tool outputs, logs, RAG chunks, files and history before they cost you tokens — reversibly, and locally. What it is, how the router-plus-compressors design works, what its self-reported benchmarks actually show, and the honest costs.
Ponytail is a one-file 'lazy senior dev' skill that makes coding agents stop and pick the simplest solution that works — cutting code by ~54% (and tokens ~22%) on a real agentic benchmark. What it is, how the ladder works, what trusted reviewers found, and the honest caveats.
Ralph is a brutally simple technique — loop the same prompt into a coding agent until the task is done. What it is, how the two main implementations (Anthropic's ralph-wiggum plugin and snarktank/ralph) actually differ, the anecdotal numbers behind the hype, and an honest critical review of where it breaks.
CodeBurn is a local TUI dashboard that tracks token usage, cost, and performance across Claude Code, Codex, Cursor and ~18 other AI coding tools. What it is, how to read the dashboard, how it compares to ccusage, and its pros and cons.
Graphify is a Claude Code skill that compiles code, docs, papers and images into a queryable knowledge graph — so an agent answers from a tiny subgraph instead of re-reading raw files. How it works, how it saves tokens, the real benchmarks, and how it compares to vector RAG and the LLM Wiki.
A Claude Code skill that makes the agent talk like a caveman — cutting output tokens by up to ~75% while keeping full technical accuracy. Why, how it works, how to install it, and real before/after numbers.