jongouveia/fable-mode

Fable Mode is a Claude Code skill that tries to capture how a frontier AI model reasons through hard tasks, so that other models such as Opus, Sonnet, or GPT can follow the same process. The idea started from watching Claude Fable 5 work: it plans carefully before acting, checks facts instead of relying on memory, challenges its own conclusions, and confirms results actually work before calling a task finished. Cheaper models are capable of these habits too, the project argues, they just are not usually prompted to use them. The skill breaks this process into five ordered steps, called gates, where each step must be completed before the next one starts. First, the model scopes the problem: it defines what success looks like, lists likely ways the plan could fail, and checks whether a simpler approach exists, before writing any plan. Second, it grounds its reasoning in evidence gathered during the current session rather than assumptions recalled from training or memory. Third, it reasons adversarially, attacking its own draft answer and then handing it to a separate review pass that only sees the result and its requirements, not the original reasoning behind it, since that reasoning could bias the review toward agreement. Fourth, it verifies the change actually works end to end against the success criterion from the first step, since a successful build alone is not proof the task is done. Fifth, it calibrates its effort to match the difficulty of the task and reports back clearly, stating what was verified compared with what was only assumed. You can install Fable Mode as a Claude Code plugin with a couple of commands, or copy the skill folder manually into your local or global skills directory. Since the instructions are written in plain markdown, other AI coding tools such as Codex and Cursor can also load and follow them. You invoke it directly for hard or high stakes problems, or let it trigger automatically when a non frontier model faces a nontrivial task. Simple one line edits are meant to skip the process entirely. The project is free, independent, and released under the MIT license.