Academic Deep Search

Academic Deep Search

Use this skill for requests such as:

• "What markers do studies on topic X usually measure?"
• "What do results sections in this field usually report?"
• "Show me a typical figure for this disease model or pathway."
• "What experimental methods are commonly used in this literature?"

The goal is not just to find papers. The goal is to read enough of the right papers to give the user a structured, directly useful answer.

Two Output Modes

Choose the mode that best matches the user request.

Body Mode

Use when the user asks about:

• molecules or markers commonly measured
• methods commonly used
• what findings usually appear in Results sections

Organize the answer by experiment type or finding category, not by paper.

Figure Mode

Use when the user asks for:

• a typical figure in a topic
• how findings are visually presented
• figure captions or representative panels

Organize the answer by figure, with source attribution and caption context.

Workflow

1. Clarify The Research Question

Identify:

• the topic or disease area
• whether the user wants methods, markers, findings, or figures
• whether the user named a specific journal, database, or URL
• whether the topic is biomedical or from another field

If the topic is biomedical, translate the idea into standard English search terms and prefer controlled vocabulary when possible.

2. Respect Source Scope First

If the user specifies a source, that scope is binding.

Examples:

• if the user says PubMed, do not mix in Google Scholar
• if the user names specific journals, search only those journals
• if the user gives a URL, read that source directly before searching elsewhere

Do not silently broaden the source list.

3. Build Search Terms Carefully

Use English search terms for database queries, even if the conversation is in Chinese.

For biomedical topics:

• prefer MeSH or other standardized vocabulary when available
• generate a few close variants or synonyms
• keep journal names exact when filtering by journal

Detailed query construction tips are in references/query-guide.md.

4. Search For Candidate Papers

Prefer the best database for the topic:

• biomedical: PubMed or PMC first
• quantitative or engineering topics: field-appropriate databases
• broad discovery: web search only when a better native source is unavailable

Aim to identify a small set of relevant papers with accessible full text. A few well-read papers are better than many shallow hits.

5. Verify Source Membership Before Citing

Before you cite a paper as belonging to a target journal or source, verify it.

Check:

• journal field on the abstract page or database result
• exact source metadata in the database
• whether the paper truly matches the user-specified scope

Do not attribute a paper to a journal or database unless you confirmed it.

6. Read Full Text Strategically

Abstract-only answers are usually not enough.

Read:

• Methods, Results, and Discussion for Body Mode
• figure captions plus the relevant Results text for Figure Mode

If full text is not available, say that clearly and lower confidence.

7. Select And Synthesize

Choose 2 to 5 papers that are:

• relevant to the question
• compliant with the requested source scope
• diverse enough to avoid overgeneralizing from one paper
• rich enough in methods, results, or figures to support the answer

Then synthesize across papers instead of writing a paper-by-paper summary unless the user asked for that.

Output Rules

• Answer directly in chat unless the user asks for a file.
• Use inline citations such as PMID, PMCID, DOI, or direct links.
• Be explicit about what was actually read.
• If evidence is limited, say so plainly.

Use references/query-guide.md for output templates.

Verification Checklist

Before finalizing:

• User-specified source scope was honored and not silently broadened.
• Query terms and database/source names can be reported if asked.
• At least the relevant Methods/Results/captions were inspected when full text was available.
• Source metadata was verified before claiming journal/database membership.
• Synthesis is organized by finding/method/figure type, not just paper-by-paper dumping.
• Evidence limits are stated when only abstracts, sparse hits, or indirect models were available.

Non-Negotiable Rules

• User-specified source scope overrides your defaults.
• Do not answer methods or marker questions from abstracts alone if full text is available.
• Do not fabricate source membership or figure details.
• Prefer PubMed and PMC for biomedical literature.
• Translate search intent into English for querying, but answer in the user's language when appropriate.

If Results Are Sparse

When little is found:

• broaden or narrow terms thoughtfully
• try synonyms or controlled vocabulary
• explain what was searched and why the yield was limited
• suggest the next best search strategy

PubMed Journal-Scoped Full-Text Download Pattern

Use this only for small journal-scoped lookups or a few PDFs while answering a literature question. If the user asks to build/import a reference library, create README/RIS/metadata, download many PDFs, or classify papers into relevance folders, switch to the dedicated pubmed-journal-reference-import skill.

Keep this pattern keyword-agnostic: use the user’s current topic terms, but do not bake those terms into the reusable workflow.

1. Treat the journal scope as binding. Query PubMed with the exact [jour] filter plus topic terms; do not silently include adjacent titles such as (N Y) variants.
2. Verify each candidate from PubMed XML before saving: MedlineJournalInfo/MedlineTA, Journal/ISOAbbreviation, title, PMID, DOI, PMCID, year, and article type. Exclude corrections unless the user explicitly wants them.
3. Prefer recent + free full text by combining date limits and free full text[sb], but consider older highly relevant free full-text papers separately and label them as supplemental rather than pretending they satisfy the recency preference.
4. For PMC PDFs, do not rely only on /articles/{PMCID}/pdf/; PMC may return an HTML “Preparing to download ...” proof-of-work page. Instead, fetch the PMC article page, parse the main pdf/*.pdf href, request it, and if a POW interstitial appears, parse POW_CHALLENGE, POW_DIFFICULTY, and POW_COOKIE_NAME, compute a SHA-256 nonce whose hex digest starts with the required zero prefix, set the cookie as {cookie_name}={challenge},{nonce}, then retry the PDF URL.
5. Verify saved PDFs by checking file magic %PDF and non-trivial file size. Keep search/download metadata in an artifact directory and write a concise README index plus RIS/BibTeX when useful.
6. Preserve active manuscript/source directories: save downloaded references under the requested reference directory, and put scripts, logs, JSON/CSV metadata, and failed-download traces under an artifact/archive directory.