farcaster-digest

<!-- autoresearch: variation B — sharper output via cluster + signal score + insight extraction; folds in A's semantic/channel inputs and C's source-status footer + persistent dedup --> > **${var}** — Topic filter or channel name (e.g. "prediction-markets", "base", "ai-agents"). If empty, uses default interest areas. If `${var}` is set, focus curation on that topic or channel. Read `memory/MEMORY.md` for current interests. Read the last 2 days of `memory/logs/` for recency context. Load persistent dedup state from `memory/state/farcaster-seen-hashes.json` (auto-created; safe to delete to reset dedup). The file starts as `{"hashes": [], "updated": null}` and stores cast hashes seen in the last 7 days — casts present here are skipped even if they re-appear in algorithmic feeds. ## Thesis Curation is the biggest lever, not fetching. Raw engagement counts rank by popularity, not conversation. This skill ranks by a signal score, clusters casts into 2–3 sub-narratives, leads with a one-line conversation-shape header, and demands an original insight per cast — not a paraphrase of the cast text. ## Steps ### 1. Fetch casts from three complementary sources For every endpoint, prefer `WebFetch` over curl (sandbox often blocks curl — see CLAUDE.md). Auth header: `x-api-key: $NEYNAR_API_KEY`. **(a) Topic search — literal + semantic.** Run each default topic query twice: once with `mode=literal` (default), once with `mode=semantic` to catch thematic matches the keyword query miss...