What is fable-soul?

Fable Soul is a set of rules meant to change how an AI coding agent, such as Claude Code or Codex, actually behaves rather than just what it knows. The README argues that when AI agents fail, it is usually not because they lack coding knowledge, but because they lack the operating discipline a senior engineer applies automatically, such as running a test before saying a bug is fixed, or naming the real mechanism behind a problem instead of just making a symptom go away. Each rule in the project was written after a specific real failure and tested to confirm it changes the agent's behavior, rather than being general best practice advice. The project is pure Markdown plus two small Python scripts, with no dependencies. Its core file holds around 19 rules organized into checkpoints called Operating Gates. Before acting, the agent is expected to be able to state the actual goal, the mechanism behind the situation, and what evidence would prove success. Verification is matched to the kind of task: a code bug needs a reproducing test and its output, a frontend change needs the page actually loaded and inspected. The rules also cover finishing reversible work instead of stopping to ask permission, leading an answer with the outcome instead of a long report, and treating a surprising result as something to investigate rather than explain away. A distinctive part of the project is a table of excuses the README says AI models commonly give for skipping these steps, such as assuming an edit that succeeded means the fix worked, paired with why each excuse is wrong. There is also a checklist the agent is meant to run before ending each turn, looking for signs like a hedge word on a claim that could be checked directly. The project offers three ways to use it: loading the rules into a session so the agent behaves accordingly without narrating them, growing and syncing the rule set over time as new failures are captured, and converting a stronger model's advice or postmortems into new durable rules using a set of provided prompt templates. It works with Claude Code and Codex, is released under the MIT license, and includes English and Traditional Chinese versions of its documentation.