token-lean

token-lean is a set of written instructions, not a program or a server, that teaches an AI orchestrator how to run a team of AI agents without wasting its own context window on bulk work. It is a single markdown file that any model or tool capable of reading markdown can use, including Claude Code, Codex CLI, Cursor, Gemini CLI, and custom agent setups.

The core idea is one rule: never generate bulk, and never absorb bulk. Instead of the main orchestrating agent reading every file itself, it should delegate that reading to cheaper, smaller agents that hand back compact summaries, while the orchestrator focuses on making decisions from those summaries. The README lays out a ladder of roles: a scout that does lookups and surveys and reports back briefly, a worker that handles well specified mechanical edits, a builder that handles harder multi file work and judgment calls, and a panel of independent agents used when a single model's answer cannot be trusted alone. It also lists eight specific practices, such as sending a scout instead of reading more than three files yourself, keeping a stable prompt prefix so caching stays cheap, and writing repeated explanations into a permanent rules file instead of repeating them.

Installation is mostly copying the instructions into the right place for your tool. For Claude Code it can be installed as a simple skill file, or as a fuller plugin that adds real dispatchable scout, worker, and adjudicator agents plus a hook that nudges the orchestrator after too many consecutive file reads. For Codex CLI, the same content can be pasted into an AGENTS.md file or installed as native equipment with profile configs. For Cursor, the same text goes into a project rule file.

The README states this discipline was originally written by an AI model documenting its own working style, then generalized to work across different model families. It is released under the MIT license.