caveman
A Claude Code skill that cuts ~65% of tokens by talking like a caveman.
About caveman
caveman is a gloriously simple Claude Code skill with a serious payoff: it strips your prompts down to terse, caveman-style phrasing — "why use many token when few token do trick" — and cuts roughly 65% of token usage as a result.
It is half meme, half real optimization. The underlying insight is sound: models do not need your politeness or filler to understand intent, and trimming it saves money on every call. Worth a look if your Claude Code bills are creeping up.
More Github
promptfoo
promptfoo lets you test your prompts, agents, and RAG systems with simple declarative configs — plus red-teaming and vulnerability scanning to catch failures before your users do. Compare GPT, Claude, Gemini, and DeepSeek side by side. It plugs into the command line and CI/CD, so evals run automatically on every change instead of living in someone’s head. Used by teams at OpenAI and Anthropic, it is one of the most practical ways to make agent quality measurable.
browser-use
browser-use makes websites accessible to AI agents — it hands your agent a real browser so it can navigate, click, fill forms, and complete tasks online instead of being limited to APIs that may not exist. Built on Playwright with a clean Python interface, it has become one of the most popular ways to give agents hands on the open web. Ideal for automating research, data entry, and any workflow that lives inside a browser.
dify
dify is an open-source platform for building agentic workflows and RAG pipelines with rich model and tool support, all in one collaborative workspace. It combines visual building with the escape hatch of custom code. You can deploy it on cloud, in a VPC, or fully self-hosted, so teams move from prototype to production without rebuilding the stack. If you are assembling agents with real tool use and retrieval, it is one of the most complete open options.
agent-skills
agent-skills is Addy Osmani’s collection of production-grade engineering skills for AI coding agents — reusable, opinionated instructions that teach your agent how a senior engineer actually approaches a task, not just how to autocomplete. It works across Claude Code, Codex, Cursor and Antigravity, so you can drop battle-tested skills into whatever harness you already use. A strong starting point if you want your agent to write code that fits real engineering standards.
chrome-devtools-mcp
chrome-devtools-mcp is an official MCP server that hands coding agents the full Chrome DevTools protocol — inspect the DOM, read console errors, trace network requests, and drive the page like a real browser session. It turns "the button doesn’t work" into something your agent can actually debug, live, instead of guessing from source. If you build web apps with an agent in the loop, this is one of the highest-leverage MCP servers to add.