A. Literature and Resource Discovery: Assembling the Search Stack

Purpose: No single tool covers the full landscape of literature discovery. This document maps the tools available — including Gemini Deep Research, Claude Code database skills, and specialised academic search tools — and shows how they chain together into a coherent discovery workflow. The conceptual framing (why finding is fragile, and why it differs from synthesizing and interpreting) is in B.lifecycle.2.literature. This document is the practical companion: what to use, when, and in what order.

The discovery problem

Claude is weak at finding literature. Training cutoff, paywall blindness, hallucination risk, and field-specific thinness mean that asking Claude to find you relevant papers is one of the highest-risk uses of AI in research. See B.lifecycle.2.literature for the full account.

The solution is not to avoid AI for discovery — it is to use the right tools for the right parts of the problem:

Part of discovery Right tool
Broad web landscape — what exists publicly on a topic Gemini Deep Research
Structured academic database search Claude Code skills (K-Dense)
Quick factual lookup with citations Perplexity
Structured synthesis tables across many papers Elicit
Citation network exploration Research Rabbit / Connected Papers
Managing and searching what you have found Zotero + Claude

Gemini Deep Research: the entry point for unknown territory

When you are entering a topic you do not know, Gemini Deep Research gives you the broadest initial picture fastest. It searches the open web across dozens of queries, reads many sources, and produces a structured report with citations.

What it is good for:

What it cannot do:

Prompt pattern:

"What is the current state of [topic] in [field]?
What are the main methodological approaches, key scholars,
and most debated questions? What has been published in the last
three years that is reshaping this area?"

The output is a starting point — a list of names, positions, and possible sources to follow up. Add what looks relevant to Zotero; verify everything before treating it as established.

For full guidance on Gemini Deep Research, see A.adjacent.gemini.

Can Gemini Deep Research connect directly to Claude Code?

This is a reasonable question — if you work in Claude Code, can you call Gemini Deep Research from within a session rather than switching applications?

The short answer: Gemini Deep Research and Claude Code are separate systems from different companies. There is no official Claude skill that invokes Gemini Deep Research. The standard workflow is sequential: run Gemini first, bring the output into your Claude Code session as a file or pasted text, then continue.

What Claude Code does have for in-session deep research:

Two skills provide analogous functionality within Claude Code itself:

Weizhena/Deep-Research-skills — five skills forming a structured human-in-the-loop research workflow: generate an outline → expand → parallel web investigation → synthesise to Markdown. Designed to keep the researcher in control at each stage. This is the closest Claude Code equivalent to Gemini Deep Research's process, but applied to more focused questions. See A.skills-ecosystem.

Imbad0202 Deep Research — 13-agent research team using the FINER scoring framework (Feasible, Interesting, Novel, Ethical, Relevant). More powerful but token-intensive; requires Max plan. Best for PhD researchers doing intensive publication preparation.

A possible technical integration via MCP: Google provides API access to its AI products, and the MCP ecosystem evolves quickly. It is worth checking whether a community-maintained MCP server for Google Search or Gemini is available by the time you read this — the MCP server registry is the place to check. If one exists, it would allow Claude Code to query Google search directly. This would not be the same as Gemini Deep Research's multi-step research process, but would give Claude Code live web search capability.

Claude Code database skills: structured academic search

For systematic literature search within Claude Code — searching actual academic databases rather than the open web — the K-Dense scientific skills collection provides direct database access:

openalex-database ★ Searches OpenAlex (240M+ scholarly works, no API key needed). Covers humanities and social sciences well — broader than PubMed. Up to 50 items per query with CSV export. Author search, citation tracking, open access filtering. The best starting point for DISSINET-relevant literature.

/openalex-database
"Find papers on social network analysis in medieval religious history,
 published 2015–2025, sorted by citation count. Export to CSV."

literature-review ★ Full systematic review workflow: PICO-framework search planning → multi-database search (PubMed, arXiv, Semantic Scholar) → deduplication → screening → thematic synthesis → PRISMA flow diagram → PDF output. The most complete skill for a thorough review. Best for medical/biological literature; use alongside openalex-database for broader coverage.

arxiv-database Searches arXiv — useful for computational social science, digital humanities, network science, and methodological literature. Keyword, author, and category search with Boolean support.

pubmed-database Medical and biological literature. Relevant for DISSINET if work intersects with medical history or health sciences. PICO-framework support; 10,000 results per search with API key.

scholar-evaluation Evaluate sources for quality, credibility, and methodological rigour. Use after a search to filter before committing sources to your synthesis.

Installation: see A.skills-ecosystem for the K-Dense installation instructions.

Elicit and specialised academic tools

For structured literature synthesis across many papers — where you want to extract specific fields (population, method, findings) systematically — Elicit is purpose-built for this. It produces evidence tables rather than prose summaries, which suits systematic review and quantitative synthesis better than Claude's default output.

Research Rabbit and Connected Papers explore citation networks — finding what a paper cites, what cites it, who else is in the intellectual neighbourhood. These are useful for discovering literature through intellectual proximity rather than keyword search; they are particularly useful once you have a few anchor papers.

For full descriptions and comparison, see A.adjacent.tools.

The discovery workflow

A practical sequence that combines the tools above:

Step 1 — Broad landscape (Gemini Deep Research or Perplexity) Run a Gemini Deep Research query on your topic. Read the report. Note the names, positions, and debates. Do not add citations to Zotero yet — verify them first.

Step 2 — Citation network (Research Rabbit / Connected Papers) Take the 2–3 most relevant papers from Step 1. Feed them to Research Rabbit or Connected Papers. Find who else is in that intellectual neighbourhood. This discovers literature that keyword search misses.

Step 3 — Systematic database search (Claude Code + openalex-database) Run structured searches in OpenAlex, arXiv, or PubMed via Claude Code skills. Use the names and terms you found in Steps 1–2 to sharpen your queries. Export to CSV.

Step 4 — Capture to Zotero Add relevant items from Steps 1–3 to Zotero. Sync. At this point, your Zotero library is the authoritative record of what you have found; everything else was exploration.

Step 5 — Synthesize in Claude Code (with Zotero MCP) Connect Claude Code to your Zotero library (A.setup.zotero-mcp). Now you can ask Claude to synthesize what you have actually collected — not what it thinks exists. The synthesis skill (or a focused session prompt) works with your verified sources, not with training data or web search.

Step 6 — Evaluate and gap-check Use the scholar-evaluation skill or the literature-review skill's screening stage to assess what you have. Ask Claude: "Given what I have found, what does my collection not cover that it should?" This is a useful gap diagnostic — Claude can identify what types of sources are missing from a set you provide, even if it cannot reliably find those sources itself.

A note on verification

At every step, citations produced by AI tools must be verified. Gemini Deep Research grounds its citations in actual URLs — lower hallucination risk than Claude — but URLs go dead, titles get paraphrased, and source quality varies. Claude's own citation generation carries the highest hallucination risk and should never be used without verification.

The rule: a source exists in your literature review when you have seen the actual paper, confirmed its title and author, and read enough to know it says what you claim it says. Not before.

Creating your own search skill

If you work repeatedly with a specific database, source type, or search protocol, consider encoding it as a skill. Examples:

Field-specific search skill. You always search the same combination of databases (e.g., JSTOR + OpenAlex + a specific regional journal index) with the same Boolean structure. Encode the protocol as a skill that runs the full combination from a single invocation.

Source evaluation skill. Your field has specific quality criteria — journal ranking systems, source type hierarchies, language requirements. Encode them so Claude applies your field's standards rather than generic criteria.

For the skill creation workflow, see A.skills-ecosystem — the "Creating your own skills" section.

Related