echo-of-machines/fable-advisor

This repository holds two matched Claude Code skills, fable-advisor and fable-coordinator, that both deal with the same idea: how a smaller AI model called Fable 5 should be used inside a coding session. The two skills are split by which model is running the session, so only the half that applies actually loads. When the session's main model is Opus, the fable-advisor skill applies. In that setup, Opus plans and directs the work, Sonnet models handle the bulk of routine tasks, and Fable is brought in specifically to advise at important decision points, the moments where it helps to check with a separate model before committing to a choice. When the session's main model is Fable itself, fable-coordinator applies instead. Here Fable takes on the coordinator role: it keeps the planning and judgment calls that only it can make, hands off self-contained pieces of frontier-level work to a separate delegate, and spreads the remaining routine tasks across Sonnet workers. On a session running Sonnet or Haiku as the main model, neither skill is meant to apply. Because a model cannot always tell what it currently is, especially right after someone switches models mid-session, both skills start with a small check called the step-0 model probe. This is a shared script that looks directly at the session's history to figure out which model is actually running right now, rather than trusting the model to describe itself. It reports back one simple label, such as opus, fable, sonnet, haiku, or unknown, and each skill uses that label to decide whether to proceed or hand off to its sibling skill. To use either skill, you copy both folders into your Claude Code skills directory, for example inside a folder named skills in your home directory, so Claude Code can find them automatically when a session starts. The two skills are meant to be installed together, since the fable-coordinator skill relies on the model-detection script that lives inside the fable-advisor folder.