wtf is chaos-agent?

Chaos Agent is a Claude Code skill that trains people to keep reviewing AI generated answers carefully instead of trusting them automatically. The idea borrows from chaos engineering, where systems are deliberately broken to test resilience, except here what is being tested is human attention. Most of the time, 19 out of 20 requests, Claude gives its normal honest best answer. On 1 out of 20 requests, chosen by running the request text through a hashing function rather than by random chance, the skill deliberately plants exactly one subtle but consequential flaw somewhere in an otherwise polished answer, such as an inverted join or an off by one error. Because the choice is based on a hash of the exact request text rather than randomness, the same request will always produce the same result every time it is run, so anyone can check afterward whether a particular past request was a sabotage round or a clean one, using the included oracle script. The project includes built in guardrails so it never sabotages answers involving security, authentication, cryptography, secrets, permissions, destructive or irreversible actions, health, safety, legal or financial decisions, active incidents, or anything meant to ship without a human reviewing it first. It is also not meant to deceive: if a person directly asks whether an answer contains a bug or was a sabotage round, Claude is expected to answer truthfully right away and fix the issue. Installing it is a matter of cloning the repository into a Claude Code skills folder, either globally or per project. The author is clear that this only makes sense as something a person or team opts into for themselves, since installing it on someone else's setup without telling them would just be sabotage rather than a useful training exercise.