research-signal-curator
Curate time-sensitive research signals for PaperTrace, including papers, model launches or retirements, tools, Skills, MCP servers, labs, events, and public community discussion. Use when adding or ra
SKILL.md.agents/skills/research-signal-curator/SKILL.md
GitHubResearch Signal Curator
Turn a noisy stream into a small, timely, auditable action queue.
Source ladder
Use social posts, LINUX DO, Hacker News, newsletters, blogs, and media to discover a lead. Verify the publishable claim against the strongest available source:
- paper, official announcement, documentation, repository, dataset, benchmark, organizer page, or status API;
- structured indexes such as arXiv, Hugging Face, OpenAlex, or a field-specific database;
- public community and social discussion for attention and practitioner pain only.
Record a public URL and verification timestamp. Do not scrape private sessions, auto-post to communities, or infer unavailable engagement.
Separate the signals
- Impact: evidence strength, technical consequence, downstream work, credible adoption, or citations.
- Buzz: recent, platform-local public attention. Never merge unlike metrics or treat them as scientific impact.
- Utility: runnable code, data, model, API, demo, setup cost, and fit to a real task.
- Attention: skim, read, or deep dive based on the next decision the reader can make.
Scores are editorial heuristics. Explain them and keep vendor benchmarks labeled. A registry upload is not automatically a launch; a service incident is not a model retirement.
Inclusion test
Publish only when the item answers all of these:
- What changed, on what date, and on which product/research surface?
- Who should care, and what action can they take before the signal goes stale?
- What primary evidence supports the smallest factual claim?
- What remains uncertain, promotional, preliminary, or unreviewed?
- Does it improve field balance, or is it another duplicate AI story?
For events and roles, require a future action boundary and official URL. For medical or safety-critical content, use discovery language and do not turn preliminary research into advice.
Failure behavior
Keep source failures visible. Preserve a last-known-good record only when it is clearly labeled stale. Do not silently replace missing data with fabricated counts or remove a previously useful source without explaining the failure.
After editing ingestion or editorial data, run the relevant sync/check command and npm run build. Update docs/PRODUCT_ROADMAP.md when a product capability materially changes.