danielarif26/understudy

Understudy is a Claude Code hook that speeds up routine prompts by having a small model on your own computer draft an answer first, before Claude ever sees the request. Claude then checks the draft and either delivers it or throws it out and answers from scratch if it is not good enough. The idea is that reasoning models like Claude spend the same expensive effort on simple tasks, like turning a list into JSON, as they do on hard problems, so Understudy hands the easy rehearsal work to a local model instead, while Claude keeps final judgment. Only certain kinds of prompts get offloaded: things like summarizing text, classifying items, converting data to JSON or a table, extracting emails or names from text, and proofreading, but only when the text is typed directly into the prompt. Anything that mentions files, paths, code, repositories, bugs, tests, URLs, or requires real reasoning goes straight to Claude untouched. A simple pattern matcher, not another AI model, decides what qualifies, so the check itself adds almost no delay. The tool is built to fail safely. If the local model is not ready, if Ollama, the software that runs local models, is down or slow, or if a timeout is reached, Understudy just skips and Claude Code works exactly as if it were not installed. Every draft and skip gets logged locally so you can review your own savings later using an included script. Setup requires Node 18 or newer, Ollama installed, and at least one local model already downloaded. The install script walks through picking a model and finishes with a single command. A few environment variables let you turn it off temporarily, choose which model to use, adjust the timeout, or point at a different Ollama server. An uninstall script removes everything cleanly. The project is a single file, written in JavaScript, and released under the MIT license, which allows free use including commercial use as long as the copyright notice stays intact.