effortmining

This project is a plugin for Claude Code, the AI coding assistant, that changes how hard the helper AI agents it spawns have to think. Normally, when Claude Code creates a helper agent to do a small job, that helper copies the same reasoning effort setting as the main session, even if the job is something simple like reading a file or reformatting a list. Thinking harder costs more in AI usage, called tokens, so this wastes resources on easy tasks.

The tool works by first classifying each subtask into a category, such as mechanical extraction, simple transformation, moderate reasoning, or hard reasoning, and then looking up the cheapest effort level that has been shown through testing to still produce correct results for that category. It then dispatches the helper agent at that specific effort level instead of the default.

Because Claude Code does not let a user set effort per individual helper directly, the project works around this by shipping five nearly identical helper agent definitions, one for each effort level, and picking which one to use based on the classification. A lookup table stores which task category maps to which effort level, and that table was built from measured benchmark runs rather than guesswork.

The project reports that in its own testing, using this calibrated approach used about 65 percent fewer output tokens than letting every helper inherit the main session's effort level, while getting correct answers just as often. It also includes tools to re-run these benchmarks and refit the lookup table for different AI models.

Installing it requires Python 3 with no extra libraries, and it is added to Claude Code as a plugin from a marketplace command. Once installed, it works automatically from a new session without further setup. This is aimed at developers already using Claude Code who want to reduce their AI usage costs.