One-prompt ArXiv filter: parse email digest, output three papers (quickchat.ai)

A developer built a compact automation that turns the noisy daily arXiv digest (~200 CS papers/day) into a personalized “3 must-reads” email using a single high-reasoning GPT-5 prompt. The pipeline fetches the arXiv digest via IMAP, extracts and cleans the email body, appends it to a carefully engineered system prompt, calls ChatOpenAI (gpt-5, reasoning_effort: high) to return 0–3 picks in a Structured Pydantic schema, and then emails the curated list via SMTP. The prompt encodes strict selection criteria (prefer 1–2 picks unless 3 are exceptional), domain priorities (LLMs, agentic systems, cognitive science, theory-of-mind, directional shifts like new memory/planning approaches), and output rules (exact title, only include canonical arXiv URL if present, concise motivation and “why”). This is significant because it demonstrates a lightweight, reproducible way to triage research at scale using LLMs while keeping human-in-the-loop curation standards—saving researchers time and surfacing conceptual or field-shaping work instead of incremental fine-tunes. Key technical notes and trade-offs: the script uses IMAP/SMTP with app passwords, Pydantic models for structured output, and LangChain-style prompt chaining; the prompt explicitly tries to reduce hallucinated links and enforces schema output, but reliance on an LLM still introduces risks (missed papers, subtle bias in selections, or occasional hallucination). The author recommends customizing the prompt, running the script on a schedule (cron/LaunchAgents), and minding API/email credentials.