fablize

fablize is a plugin for Claude Code (the AI coding assistant) that installs a set of behavioral rules designed to make Claude's Opus model follow through on tasks more consistently. The name refers to Fable, a newer and more capable Claude model, and the plugin's premise is that some of what makes Fable better is not raw capability but procedure: actually running code before declaring it done, checking results, and investigating problems in a methodical way rather than guessing.

The author ran a comparison between Fable 5 and Opus 4.8 across about 45 sessions and 1,500 tool calls. The conclusion was that on concrete tasks like writing and running code, the two models performed similarly. The difference appeared on more open-ended problems where Fable would follow an implication one step further without being asked. That kind of depth, the author argues, cannot be taught to a weaker model through instructions. But the working habits that support good output, such as verifying that built artifacts actually run, refusing to say a task is done without evidence, and tracing bugs to their root cause rather than guessing, can be transferred.

The plugin installs hooks into Claude Code that enforce those specific procedures. When you ask Claude to do something, a routing hook checks what kind of task it is and applies the matching discipline. For tasks involving rendered output like HTML, SVG, or charts, it requires Claude to actually open and inspect the result. For multi-part tasks, it decomposes the work and blocks completion unless each part is verified. For debugging, it enforces a structured investigation rather than trial-and-error fixes. A separate hook catches when Claude says it will do something but stops short of doing it.

The README is explicit about what the plugin cannot do: it will not improve Claude's ability to notice out-of-scope problems, add creative depth, or pursue ideas independently. For those tasks it tells you to use a stronger model or a person.