On 26 July one rule went above the rest in everything we build for our Shopify agent: no claim without a mechanism. Every sentence the agent says about an action has to come from a function's result, checked on the server.
The reason is simple. A prompt is a request, and prompt injection has no reliable patch. A rule the model can skip isn't a rule.
The commit that added it put the idea in one line: "a prompt is not a contract."
The habit is older than the rule. The second commit in the repo, on day one in March, was called "Security hardening across 31 files."
The rule, and the walls we built from it: x.com/con_golias/status/2121352219635159234
