The Daily Digest pulls from about 35 sources each morning and collects somewhere between 400 and 800 raw items. Every day I need to pick 20 of them to publish. The question I had to answer early was how to structure the AI call that does that selection.
The obvious approach is per-item scoring: send each item to a model and ask whether it belongs in the digest, given a description of the reader. Aggregate the scores, pick the top 20. Simple to reason about, easy to test in isolation.
The problem is that relevance isn't absolute. A paper on LLM red-teaming might be a clear pick if it's the only one this week, and a clear skip if three similar papers landed in the same batch. An AI lab announcement matters more when no other lab said the same thing yesterday. Per-item scoring can't see any of that. Each call is blind to the rest of the batch.
Per-item also has an economics problem: somewhere between 400 and 800 API calls per run, every single morning. That's both slow and expensive for what is, in the end, a morning briefing.
The design I landed on is a two-step process. A pre-filter first: sort each source category by recency and cap it at 18 candidates. This collapses 700-plus raw items to roughly 100. Then one Claude call sees all 100 candidates and picks and annotates the final 20 in a single pass.
That pre-filter step had a non-obvious constraint. Early tests sent all raw items (sometimes 700-plus) in a single call without any cap. The context limit wasn't the bottleneck. Attention degradation was. Picks from the bottom of a 700-item JSON array weren't as sharp as picks from a 100-item one.
Capping at 18 per category is a heuristic, not a derived number. The underlying principle: reduce the candidate pool to something the model can reason across coherently before asking it to rank. That constraint made a visible difference in output quality.
The single-call approach does something per-item scoring can't: catch relationships between items. The curation prompt asks the model to flag connections between selected items: same threat actor, same technology, same underlying event. That cross-item annotation only works when everything is in context at once.
One call per day. The whole run takes about 90 seconds from cron fire to published post.