deep-search — Claude Code skill: research search and install
The deep-search Claude Code skill routes research searches through web search, literature APIs, or relation counts. It helps investigate thin, empty, or unusable results.
MIT ★ 93 updated 2026-11-15
npx skills add Insilico-org/plsi --skill deep-search -a claude-codeWhat deep-search does
deep-search chooses a route for finding research evidence and helps explain why a search returned nothing useful. It covers web search, tools for Europe PMC, PubMed, CT.gov, OpenFDA, and Evipedia, plus plsi research discover --about "<subject>", which counts links in PubTator annotations.
The skill separates a genuine lack of papers from failed retrieval and tool limits, such as evidence held in a table, figure, or full text. It calls for trying another route before assigning one of those explanations. Relation counts come from a model rather than curation, and the index covers PubMed and its open-access subset.
Queries need to fit each source: PubMed joins every token with AND, while Europe PMC lists newest records first. Fetching one record there requires EXT_ID:<pmid>. When resolved: false, PubTator uses another entity name; pass its identifier with --entity or switch to web search.
What it helps with
- find papers when a straightforward query returns few results
- choose a source for a research question
- distinguish missing evidence from failed retrieval
- spot open research questions through PubTator relation counts
Example prompts for the agent
- Search for papers on this topic; the first query returned little, so try another source.
- Explain why the search found many items but no usable evidence.
- Check PubTator links for this subject and separate well-studied questions from open ones.
How to install deep-search
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill deep-search -a claude-code2Manually: put the files in place
~/.claude/skills/deep-search/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/deep-search/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill deep-search -a codex2Manually: put the files in place
.agents/skills/deep-search/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill deep-search -a cursor2Manually: put the files in place
.cursor/skills/deep-search/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Insilico-org/plsi skills/lab/deep-search2Manually: put the files in place
.github/skills/deep-search/SKILL.md1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill deep-search -a gemini-cli2Platform CLI
gemini skills install https://github.com/Insilico-org/plsi --path skills/lab/deep-searchManually: put the files in place
.devin/skills/deep-search/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about deep-search
How do I install deep-search in Claude Code?
Run npx skills add Insilico-org/plsi --skill deep-search -a claude-code. The skill file is at ~/.claude/skills/deep-search/SKILL.md.
What does deep-search do?
It selects a route for research evidence searches and helps distinguish absent results, failed retrieval, and tool limits.
How does deep-search diagnose an empty search?
It suggests trying a different route. Searches across different vocabularies that still return nothing can indicate a lack of papers; results with no usable evidence point to retrieval failure.
Source text
SKILL.md · MIT
--- name: deep-search description: Use as PLSI Research Orchestrator when searching for evidence in research and the obvious query is not enough - choosing between your own web search, the literature APIs and the relation counts, adapting a query per source, and telling a silent field from a failing instrument. Load it from the Gather move when a search comes back thin, junk or empty, and from Ask when a direction needs to be aimed before it is searched. --- # Finding what is actually there `lab/gather` says how to search and what a finding must carry; this says which route to take and what a search returning nothing means. ## Three routes - **Your own web search** reaches the whole page: results tables, registry entries, preprints, guidelines, the paper a relevance ranker will not return. Use it for the seed, for numbers not in any abstract, and to check whether an empty API result is the field or the query. - **The toolbelt** gives verifiable identifiers and publication types. Only `europepmc` returns abstracts; `pubmed` for publication types and hit counts, `ctgov` trials, `openfda` labels, `evipedia` evidence reviews. - **`plsi research discover --about "<subject>"`** reads PubTator relation annotations and returns, for each entity linked to the subject, how many papers stand behind the link. It tells a settled question (`well_studied`) from an open one (`candidates`, each with its PMIDs as seeds) before you spend a cycle. The counts are a model's reading, not curation, so the quote still comes from the paper. It indexes PubMed and the open-access subset only. `resolved: false` means PubTator uses another name for the subject: pass its id with `--entity` (`@CHEMICAL_<Name>`, `@GENE_<SYMBOL>`, `@DISEASE_<Name>`) or use web search. Neither the APIs nor web search alone is a search. ## Queries The seed paper, the concept blocks, the five query families, the hit count, "none found" and stopping are `lab/gather` steps 1-4 and 6. Adapt each query per source: PubMed ANDs every token, so a natural-language question returns zero; Europe PMC returns the newest first and needs `EXT_ID:<pmid>` to fetch one record. ## Three different silences - **Nothing retrieved, twice, in different vocabularies**: a genuine absence, `unstudied`. - **Plenty retrieved, nothing usable**: retrieval failed, not the field, `nothing-usable`. Say what came back. - **The result exists where this loop cannot read it** (a results table, a figure, a full text): a tooling limit. Name it as its own gap. Before writing any of the three, try the other route.
© Insilico-org · MIT · Insilico-org/plsi