Splitting the pipeline: Hermes gathers, Claude decides

The first real design decision on this project was not which model to use but where to draw the line between two of them.

Everything upstream of a verdict is gathering: following sources, pulling the last day of posts, throwing away the ninety percent with no reproducible workflow in it. That work is embarrassingly parallel, each unit is disposable, and a failure just means retrying one source. Short-lived Hermes workers, one per source, fit it exactly.

Everything downstream is judgement: does this actually improve the step it would replace, and what is the tradeoff. That needs the project context, the current workflow pages, and write access to ClickUp and WordPress — all of which Claude already has in this workspace. Running it as scheduled Claude tasks removes a layer of glue code that would otherwise exist only to move context around.

The handoff is a queue: watchers write candidate entries as drafts, the Organizer drains them each morning. If Hermes scheduling turns out to be awkward, the fallback is Claude running the watchers too. That would cost parallelism and money, not capability — which is the right kind of fallback to have.