本文へスキップ
TakoAPI
スキル一覧に戻る

arxiv-search-collector

xukp20 作

Git & GitHub
859ダウンロード数はclawskills.shが公開している値です
0閲覧数
0件のいいねサインインするとこのスキルにいいねできます
—まだ評価がありません
あなたの評価

概要

Model-driven arXiv retrieval workflow for building a paper set with a manual language parameter: initialize a run.

説明

このスキルでできること

A four-step workflow for collecting arXiv papers where the model designs search queries, fetches metadata via the arXiv API, filters results by relevance, and merges selected papers into a structured directory. Query planning and relevance decisions are made by the model, not hardcoded rules.

Rule-based keyword filters often miss contextually relevant papers or include noise; model-led filtering adapts to nuanced relevance criteria that static heuristics cannot express.

使いどころ

  • Building a curated paper set for a literature review on a specific AI topic
  • Tracking recent arXiv submissions across multiple related research areas
  • Gathering papers in a non-English language using the language parameter
  • Iteratively expanding a paper collection when initial results are sparse
  • Preparing a deduplicated paper index as input for downstream summarization

ワークフロー例

入力

User asks: Collect 5-10 recent papers on LLM applications in Lean 4 formalization from the past 30 days

エージェント
  1. 1

    Run init_collection_run.py with the topic, keywords, categories cs.AI and cs.LO, target range 5-10, and lookback 30d

  2. 2

    Design 3 focused queries using OR within synonym groups and AND across semantic groups

  3. 3

    Run fetch_queries_batch.py with a query_plan.json to retrieve metadata for each query serially

  4. 4

    Read each query result list and select keep indexes based on relevance to the topic

  5. 5

    Run merge_selected_papers.py with the keep indexes to deduplicate and write per-paper metadata directories

出力

A run directory containing per-paper metadata.json and metadata.md files, plus papers_index.json and papers_index.md summarizing the curated set

インストール

clawhub install xukp20/arxiv-search-collector

CLIコマンドを実行するか、Claude Code、Cursor、Codexなどお使いのコーディングエージェントにプロンプトを貼り付けてください。

リンク