One file per skill · runs as a prompt or a checklist
Research skills
The judgment behind this atlas, written down. Each skill is a short, self-contained procedure — concrete search queries, decision rules, output formats — that a person can follow or an agent can execute verbatim. Together they form the daily maintenance loop.
discover-wam-papers
The daily sweep: concrete arXiv API queries, listing pages, and Hugging Face daily papers — with dedup rules against the existing dataset. Run by the daily discovery agent.
triage-wam-paper
Decide inclusion in under five minutes. The exclusivity bar, the five-pathway decision tree, and when to reject or escalate to Discussion.
read-wam-paper
A structured deep-read: what to extract (world-model component, action component, training data, evaluation, limitations) and the order to read sections in.
summarize-wam-paper
Turn a deep-read into a schema-valid entry for data/pending/ — field-by-field rules, including how to write one_line vs significance and pick builds_on.
verify-wam-entry
The adversarial fact-check before merge: links resolve, metadata matches the arXiv record, the classification survives a second opinion. Run by the review agent.
compare-wam-approaches
Compare two to four papers along fixed axes — representation, action interface, training signal, data scale, results — into a table and a verdict.