What is ultracode-token-optimization?

This project is a set of operating rules, packaged as a Claude Code skill, for writing token efficient multi agent orchestration scripts using Claude Code's Ultracode workflow feature. That feature lets you write scripts using functions like agent, parallel, pipeline, and phase to run several AI agents together on one task, and this skill is meant to be loaded before writing or fixing such a script so the rules get applied from the start. The rules were pulled from real experience running multiple projects at once with one agent acting as an orchestrator, another as an executor, and separate reviewer agents providing a different perspective. The author is clear that the actual target of optimization is not the raw price of tokens, but the wasted rework that happens when a task gets sent back for revision, fixed, and reviewed again and again. That kind of rework tends to come from agents overstepping their assigned role, dumping raw unfiltered logs instead of summaries, marking work as complete based only on self reported claims, or letting a reviewer agent make final decisions it should not be making. The skill writes these failure patterns down as explicit rules so they can be checked before a script is even written. The package contains one file, SKILL.md, which includes the trigger keywords Claude Code watches for, a real accounting of the Ultracode workflow runtime's actual constraints, a table showing which model should be assigned to which role such as orchestrator, executor, or reviewer, an alternative approach for getting a genuinely different reviewing perspective, sixteen firm rules for writing workflow scripts, a checklist to run through before submitting a script, and an honest note that this skill only nudges behavior in a softer way and cannot guarantee a hard limit on token usage. You install it by placing the file in your Claude Code skills folder. The project is released under the MIT license.