What is fable-token-saving-skills-orchestrator?

Fable Token-Saving Skills Orchestrator is a Python toolkit meant to be layered on top of an existing Claude Code setup, aimed at reducing how much of the expensive Fable model is used for routine work. The core idea is that Fable is well suited to judgment tasks like strategy, architecture decisions, and reviewing work, but wasteful when it is left typing routine code, rereading large logs, or waking up repeatedly from a cold context after sitting idle. The kit adds a block of instructions to a project's CLAUDE.md file, along with hooks, skill templates, and documentation, without replacing anything already in place. Hooks act as safety guards for common mistakes that happen when a long-running agent session stops. Skill templates cover things like routing work to other tools, dispatching tasks to Codex, polishing a README, and tracking the operational status of a public repository. Everything installs through a script that only makes changes when you explicitly pass an apply flag, and it writes timestamped backups before touching any existing file. The reasoning behind the approach rests on how Anthropic's prompt caching pricing works: a cache write costs more than a plain request, but reading from a warm cache is much cheaper, and the cache expires after a short idle period by default. Based on that, the project's guidance is to keep the expensive model focused on judgment and integration work, hand off routine implementation and large-output digestion to cheaper tools, use a brief keepalive read only when a wait is short enough to be worth preserving the cache, and simply let the cache expire during genuinely long idle stretches. The project is licensed under MIT and includes documentation covering installation, routing strategy, keepalive economics, and how to adapt the setup to different models. It explicitly warns against pasting real credentials into any of its scripts or examples, recommending environment variables or a separate secret manager instead.