mrgediao/paper-reading-zh

paper-reading-zh is a collection of prompt rules designed to make AI systems read academic papers more carefully and honestly, with a focus on Chinese-language workflows. The core idea is that when an AI reads a paper, it should mark anything it cannot verify as unverified rather than filling in gaps with plausible-sounding guesses. Venue names, publication years, conference rankings, and code links must be confirmed before being stated. If experimental numbers cannot be traced back to a specific table or figure in the paper, the AI must say so. The rules work in two ways. You can paste the provided prompt text into the system instructions of a Claude Project or ChatGPT Project, then upload a PDF or paste a link to a paper and ask the AI to read it. Alternatively, for command-line tools like Codex or Claude Code, you can copy the folder into the tool's skills directory and the rules activate automatically when you provide a paper and ask for a close reading. There are three reading modes. The deep-read mode walks through a single paper covering the problem it addresses, its main contributions, the method it proposes, experimental results, and limitations. The engineering mode focuses on whether the described approach could realistically be implemented: it maps out data flow, module responsibilities, and gaps in implementation detail. The survey mode handles comparisons across multiple papers, checking that datasets, evaluation metrics, model sizes, and training conditions are actually comparable before drawing conclusions. The project is in early development at version 0.1.2. Known limitations include incomplete end-to-end testing across all supported platforms, PDF chart reading ability that varies by platform, and formula-to-code alignment that only activates when the user supplies code directly. The project does not generate slides, bibliography files, or full translations.