zealchou/fable5-scaffold

Fable Scaffold is a set of shell hooks for Claude Code, the AI coding assistant, designed to add structured discipline to how the AI works through a task. The README is bilingual in Chinese and English. The core idea is that reliable AI outputs come partly from the model verifying its own reasoning before answering. This scaffold tries to replicate that behavior externally by injecting prompts and checks at specific moments in a session, rather than hoping the model applies that discipline on its own. The README is explicit that this does not improve the underlying model's intelligence, it only enforces process. A Sonnet model using this scaffold is still that model, just nudged to slow down and check itself. The scaffold divides a working session into six stages. At startup, it loads project memory and recent lessons. Before the AI gives a complex answer, it inserts a forced three-round verification: a draft, a check where the model must look up evidence for each claim, and a third pass that challenges whether the earlier reasoning holds. Before the AI creates a new file or tool, it checks whether something equivalent already exists. Before committing code, it runs a verification step. When the session ends, it prompts the AI to record what was learned. The hooks work through Claude Code's built-in hook system, which runs scripts at defined lifecycle points like session start, before a tool runs, and at stop. The hooks are advisory rather than blocking: they inject a prompt reminder but do not prevent the AI from proceeding. The stated tradeoff is that this makes responses slower and uses more tokens, so the README suggests disabling it for simple questions and using it only for complex multi-step tasks. A hotfix note at the top warns that version 1.0.0 had a false-positive bug in one hook: any command that included the literal text "git commit" as a search string would incorrectly trigger a commit check. Version 1.0.1 fixed this. The repository is not affiliated with Anthropic.