explore-sota
Use when building or expanding the SOTA — from an idea, keywords, seed papers, or the existing collection — through MCP search, citation chasing, triage, and digestion of accepted papers.
How do I install this agent skill?
npx skills add https://github.com/vincenzoimp/academic-research-skills --skill explore-sotaIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The skill automates research paper discovery and organization using scholarly APIs via MCP. It involves running a local validation command and processing external metadata, which introduces standard minor security surfaces for a research tool.
- Socketpass
No alerts
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
Explore SOTA
Run the SOTA exploration loop: discover candidate papers, triage them in
sota/queue.md, digest the accepted ones, and repeat until the declared
stopping rule is met. Works for long autonomous sessions and for small
targeted expansions alike.
Read First
sota/README.md— digestion rules and formatssota/queue.md— the Scope block and the current frontierreferences/citation-chasing.md— frontier discipline, anti-echo-chamber rules, review scales
MCP Preflight (hard gate)
Same gate as digest-paper: by capability, never by API key. arxiv must
respond AND at least one of semantic-scholar, dblp, or openalex. If the
gate is unmet: STOP and report. A missing key only throttles; a reachable
source being down degrades the cross-check, not a stop. No fallback to model
memory or web scraping.
Procedure
- Scope. Read the Scope block at the top of
sota/queue.md. If empty or stale, write it now: research question, keywords/synonyms/adjacent terms, inclusion and exclusion criteria, review scale (quick-scan ~8–15 papers / focused-sota ~20–40 / full-survey 50+), stopping rule. In an interactive session confirm it with the user; in an autonomous session derive it fromREADME.mdand record it before searching. - Seeds. Build a diversified seed set: user-named papers, already digested papers, seminal works found via MCP search, recent frontier papers. Never start from a single author group, venue, or survey.
- Search. Run keyword queries on every configured scholarly MCP — the always-on arxiv, semantic-scholar, dblp, and the paper-search aggregator, plus openalex when enabled. Short, high-signal queries; record productive terms in the Scope block.
- Chase. For each digested seed, fetch outgoing references and
incoming citations via semantic-scholar — one hop at a time, per
references/citation-chasing.md. - Dedupe. Before adding a candidate to the queue, check it is not in
sota/index.mdor already insota/queue.md(match DOI, arXiv id, S2 id, then title + first author + year). - Triage. Add each new candidate to
queue.mdwith provenance ("found via") and decide against the criteria:accepted,rejected: <reason>, orpending. A candidate no MCP can resolve getsunresolvable-via-mcpand is never cited. - Digest. For each
acceptedcandidate, run the digest-paper skill procedure (all-or-nothing). New digests yield new citation leads — feed them back into the queue. - Learn. After each digestion round, add newly learned terminology, benchmarks, venues, and author groups to the scope terms and re-search.
- Stop. Before declaring saturation, run the anti-echo-chamber checks
in
references/citation-chasing.md. Stop at saturation (a hop yields mostly duplicates or out-of-scope work) or at the declared budget — in that case record the unexpanded frontier aspendingrows, never silently. - Run
make checkand leave no untriaged row before ending the session.
Rules
- Every candidate enters
queue.mdbefore any digestion decision; the queue is the only frontier record. - Rejections always carry a reason tied to the exclusion criteria.
- Citation counts and graph centrality are discovery signals, not relevance or quality judgments.
- Never pad toward the paper budget: scale targets are budgets, not goals.
Done When
- No
pendingrow remains, or the unexpanded frontier is explicitly recorded and reported to the user - The Scope block reflects the final criteria and the saturation/budget outcome
make checkpasses
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/vincenzoimp/academic-research-skills/explore-sota">View explore-sota on skillZs</a>