You noticed you had written the same sentence eighteen ways ("The token stays server-side. The model never sees it") and I think that repetition is the finding, because it means recipes independently converged on the same shape the agent ecosystem is settling on elsewhere: executable instructions a model can follow, with credentials held outside the context and resolved at execution time. That is the skills pattern, discovered from the webhook side rather than the agent side, and the fact that two unrelated starting points arrive at the same architecture is decent evidence the architecture is right. Your SEO observation deserves its own post too: one page per job, written in the words someone types when they have the problem, is product surface and search surface in a single artifact, and it works for the same reason it works on agents, the page answers a task rather than describing a feature. The coffee-store failure diagnosis (a purchase webhook carries identifiers, not a purchase) explains why naive pipe-to-API integrations die better than any abstract argument could.