Paper Scout

A living collection of paper-conditioned research reports. Paper Scout is an agent-native workflow that scouts recent papers (default source: Hugging Face), filters hard, investigates the most promising or interesting ones deeply, and delivers a fresh Feishu report for each run. The record accumulates here as runs happen.

There is no code to build and nothing to install. The repository is the instance. What it ships is prompts — instruction files and skills that are easy to read and tune.

How a Run Works

1. scout.sh starts the reading agent from workspace/ with a date-stamped prompt.txt.
2. The agent reads workspace/AGENTS.md (its contract) and loads skills from workspace/.agents/skills/ as needed: scout the recent pool, filter aggressively, deep-dive a small handful, compose the report, deliver a fresh Feishu doc.
3. The report and preserved evidence are saved under workspace/runs/<run-id>/, and the run is logged to workspace/runs/INDEX.md.

Layout

• prompt.txt — the run trigger
• scout.sh — launches a run
• workspace/AGENTS.md — the reading agent's contract (your interests, cadence, Feishu destination)
• workspace/.agents/skills/ — the scouting, deep-dive, and Feishu-doc skills
• workspace/code/ — ignored lab bench for cloned repos, probes, patches, and temporary environments
• workspace/runs/ — delivered reports, preserved evidence packets, and INDEX.md
• AGENTS.md — contract for coding agents maintaining this repo

workspace/papers/ is the tracked paper-text cache. workspace/runs/ holds durable run packets. workspace/code/ and workspace/drafts/ hold scratch run data and are gitignored except for their README markers.

Make It Your Own

Want your own tracker? Clone this, then tell your agent something like: "read this repo and configure a paper tracker for me in the same spirit, tailored to my interests." The agent edits workspace/AGENTS.md (interests, exclusions, cadence, Feishu destination) and tunes the skills as needed. No setup script required — it's all prompts.