Autoresearch

Autoresearch is an end-to-end research agent that fans out parallel queries across the open web, ArXiv, and Semantic Scholar, then deduplicates, cross-references, and folds everything into a single citation-linked brief. It's built for people who do literature reviews, competitive teardowns, and due-diligence — work where the bottleneck isn't reading, it's collecting and reconciling dozens of sources without losing the thread.

When to use it

• Scoping a literature review and you need the seminal papers plus the last 12 months of preprints in one pass.
• Competitive analysis where claims live across vendor pages, blog posts, and benchmark papers that contradict each other.
• Due-diligence on a company, technology, or claim, where you need every assertion traceable back to a source.
• Any "I'd lose half a day in 30 browser tabs" task.

How it works

Autoresearch runs three retrieval lanes concurrently: a general web search, the ArXiv API (http://export.arxiv.org/api/query), and the Semantic Scholar Graph API (/graph/v1/paper/search, with fields=title,abstract,year,citationCount,externalIds). Results are normalized to a common record, then deduplicated by DOI / ArXiv ID first and fuzzy title match second, so the same paper surfacing on the web and in ArXiv collapses to one entry.

It then does a cross-reference pass: each factual claim in the draft is attached to the source(s) that support it, and conflicting claims are flagged rather than silently averaged. The final output is a Markdown brief with inline numbered citations and a reference list carrying URLs, DOIs, and citation counts.

Example

curl -fsSL --create-dirs -o ~/.claude/skills/autoresearch/SKILL.md https://claudeskil.com/skills/autoresearch/SKILL.md

# Then, in a Claude session:
> Use autoresearch: "retrieval-augmented generation for long-context
  models" — focus on 2023 onward, prioritize peer-reviewed and ArXiv,
  output a brief with a contradictions section.

The skill returns brief.md (synthesis + inline citations) and a deduplicated source table you can audit line by line.

Tips & gotchas

• Scope the query tightly. Broad prompts ("AI safety") pull hundreds of candidates and dilute synthesis. Add a year window, a sub-topic, and a source-type preference.
• Semantic Scholar rate-limits unauthenticated traffic. For large runs set SEMANTIC_SCHOLAR_API_KEY to avoid throttling and dropped results.
• Citation count is recency-biased. Strong 2024–2025 preprints will show low counts; don't let the agent rank them out — ask it to weight venue and relevance too.
• Always spot-check the contradictions section. Cross-referencing flags disagreements but does not adjudicate them.

FAQ

Does it read full PDFs or just abstracts? Ranking and dedup run on titles and abstracts; for sources where full text is reachable it pulls the body to ground specific claims. Paywalled full text falls back to the abstract.

Can I restrict it to peer-reviewed sources only? Yes — say so in the prompt and it drops the open-web lane, keeping ArXiv and Semantic Scholar.

Will the citations be real? Every reference resolves to a fetched URL, DOI, or ArXiv ID. If a claim can't be tied to a retrieved source, it's marked unverified rather than cited.