llm-paper-radar

Use when the user names a whole conference/venue (e.g. "MLSys 2026", "分析下 NeurIPS 2026 的推理优化论文", "做一份 ICML venue 趋势报告", "venue trend report for <conf>") and wants its accepted papers fetched, filtered to LLM inference deployment optimization, grouped by subfield, and turned into a trend report with a macro synthesis. NOT for a single paper (use paper-interpret) and NOT for the daily rolling digest (that's scripts/daily.sh / paper-triage).

Use when the user gives a single paper (arXiv id, arXiv URL, or paper name/acronym) and wants it interpreted — 翻译成中文 / 解读 / 讲原理 / 溯源 / paper-river / "帮我看这篇 paper" / "interpret this paper" / "这篇讲什么". Resolves the paper against the llm-paper-radar archive first (reusing its bilingual summary, scoring, triage history, and any existing lineage analysis), then produces the interpretation from whichever angles the user wants. NOT for triaging/accepting/rejecting a digest (use paper-triage) and NOT for fetching/re-scoring the pipeline (that's scripts/daily.sh).

Use when the user asks to triage / curate / accept-reject / ingest papers from today's or recent llm-paper-radar digests. Trigger phrases include "triage papers", "triage today's digest", "看今天的 paper digest", "/paper-triage", "/triage-papers", "scout compression papers", or any reference to walking through the radar's daily/weekly digest interactively. The skill does NOT fetch or score papers — that's the cron-driven daily.sh. It only handles the human-in-the-loop accept/reject step on top of the already-rendered digest, calling scripts/seed_add.py and scripts/seed_reject.py on behalf of the user and optionally handing accepted papers to quark-wiki.