Researchers face dozens of new papers every day, and what really drains the energy is usually not reading but filtering. Most existing tools focus on close reading of a single paper and offer no filtering or categorization; meanwhile, many seemingly valuable works turn out to be hard to reproduce because their code repositories are missing or incomplete, which makes for an expensive sunk cost. paper_reader is an automation tool aimed exactly at this pain point: it takes reproducibility as its core filtering criterion, combining a rule-based pre-filter with semantic judgment from an Agent Skill running on Kimi CLI. Every day it fetches HuggingFace Daily Papers, tags and categorizes them, and generates Chinese summaries, then relies on macOS launchd scheduling and a pure static page for unattended updates and display. This post records its design motivation, key trade-offs, and engineering implementation.
