What is fable-5?

fable-5 packages a structured working method for AI coding agents, meant for hard, multi-step tasks, and ships it as an installable plugin for Claude Code as well as several other AI coding tools. The core idea, called the Fable 5 loop, breaks a task down along the points where it can actually be checked, tackles the biggest unknown first, verifies each step with the smallest test that would fail if the work were wrong, and picks the next action based on what would actually change the plan. The repository includes a skill folder with an eight step version of this loop, a catalog of small verification checks organized by task type, common ways to break work into pieces, and a pre completion check script. It also ships two supporting agents: fable-scout, which does a read only pass to understand a codebase before acting, and fable-refuter, an adversarial checker that defaults to treating a claim as unverified whenever it is not sure. Both are registered automatically when the plugin is installed in Claude Code, and can also be used as plain prompt templates in other tools. Installation differs by tool. In Claude Code, it is added as a plugin marketplace entry and installed with a couple of commands. For Codex, Cursor, and Kimi, the repository includes matching plugin manifest folders, or a user can simply symlink the skill folder into the right location. Gemini CLI can install it as an extension, or pick it up automatically if the skill folder is placed where Gemini scans for skills. For Antigravity, which has no plugin format, the same symlink approach applies. The README includes a small benchmark comparing three setups: a vanilla Opus 4.8 model, Opus 4.8 combined with the fable-5 loop as a system prompt, and Fable 5 running on its own, tested against deliberately tricky tasks designed so that fixing only the visible symptom would fail. All three setups passed every task in the test, so the author is candid that this particular benchmark does not show the loop clearly outperforming a plain capable model, only that adding the loop costs a small amount of extra time and causes no regressions. The README suggests the loop's real value would likely show up in longer, more ambiguous, multi-step sessions that a short benchmark cannot capture, and invites others to contribute harder test tasks.