What is value-for-fable?

Value-for-Fable (VFF) is a Claude Code plugin, written and documented mostly in Korean, that tries to make the cheaper Claude Sonnet model respond more like the more expensive Fable or Opus models. The core argument is that the quality gap between models is partly a matter of behavior patterns, not just raw capability. VFF injects a structured set of rules (derived from observing how Fable 5 actually operates) into every Sonnet response, narrowing that gap without changing the model itself. The author ran a blind benchmark comparing bare Sonnet, Sonnet with VFF, and bare Opus on a set of diagnostic and advice tasks. The result the README reports is that Sonnet with VFF scored 87.1 on a neutral 0-100 rubric while bare Opus scored 86.2 and 89.4 across two separate evaluations, putting them within noise range of each other. The output cost of Sonnet is roughly 30% of Opus, so the project claims about three times the cost efficiency for similar quality on the tested task types. The benchmark materials are included in the repository for review. VFF operates through four components. The skill file is manually triggered at the start of a session and activates a mode that changes how Claude structures its replies. The agent file is a second Sonnet instance that reviews draft responses against four fixed criteria: missing requirements, factual errors, unexplained conclusions, and exceeding the requested scope. The output style file is a passive, always-on version of the same rules that activates through Claude Code's config panel rather than a typed trigger. The hook is a shell script that injects a reminder when a session grows long enough that the original instructions might have drifted out of context. The eight behavioral rules VFF enforces cover: putting the conclusion in the first sentence, prioritizing measurement before prescription, running tool calls in parallel, verifying before declaring done, staying within the scope of the request, matching word count precisely rather than capping it, avoiding empty praise, and reporting only changed lines rather than full file dumps. Installation is available as a Claude Code plugin with two commands, or by copying individual files manually into the appropriate Claude config directories.