RW Literature Discovery

Startup

1. Read

   references/standalone.md

   ,

   references/method.md

   , and

   references/standards.md

   .
2. Read the materials provided by the user in this round; if no materials are provided, only complete the parts permitted by current evidence.
3. Retrieve

   references/atoms.jsonl

   when rule-based judgments are required; read

   references/cases.md

   when encountering similar tasks.
4. Read

   references/domain-guide.md

   when selecting methods or tools, and use

   assets/worksheet.md

   to organize deliverables.
5. Read

   references/acceptance.md

   .

   references/behavior-tests.json

   is only used for testing and not as factual information for user tasks.
6. When current literature, APIs, reporting standards, or journal requirements may change, open the official links in

   references/source-map.md

   for verification and record the date.

Work Stages

1. Rewrite the request into a specific judgment that needs to be supported, refuted, defined, or explained.
2. Establish concept blocks, synonyms, spelling variants, controlled terms, and exclusion terms.
3. Select at least two complementary sources by discipline, and record the database and retrieval date.
4. First find seed literature, then conduct forward, backward, and similar literature tracking.
5. Verify title, author, year, source, identifier, literature type, and access level item by item.
6. Deduplicate by identifier, standardized title, version relationship, and retraction status; do not treat preprints and journal versions as two separate pieces of evidence.
7. Record inclusion purposes, exclusion reasons, evidence conflicts, failed sources, and gaps for the next round of search.

Operational Rules

• Literature discovery serves judgment; the number of documents is not used as the completion standard.
• Exploratory retrieval and reproducible systematic retrieval must be recorded separately.
• The absence of results in a single database cannot be interpreted as the absence of research.
• Similar titles cannot replace verification of DOI, PMID, or publication metadata.
• Abstracts can only support content explicitly reported in the abstract.
• Distinguish the status of preprints, conference abstracts, journal articles, and retraction records.
• Citation counts can help find entry points but cannot directly prove research quality.
• Save search queries as-is, along with the date, platform, and number of returns.
• Synonym expansion must be tested alongside result noise.
• When encountering abnormal results, search for alternative explanations and contradictory evidence.
• Have a manual retrieval fallback when APIs fail, are rate-limited, or have insufficient coverage.
• Each item in the official literature list must explain which judgment it supports, limits, or changes.
• Mark each record as full-text verified, abstract-only, metadata-only, or unavailable; the supportable scope cannot exceed the obtained content.
• Search failures, permission restrictions, timeouts, and missing fields must be retained in the run records and cannot be silently deleted.
• Use interpretable criteria for literature screening; do not use journal names, author titles, or fixed composite scores as substitutes for relevance and evidence judgments.
• When determining whether data or literature is the currently available version, record the data year, publication year, verification date, expected refresh time, and publishing source.
• Only use "latest" after verifying the update time of similar sources; otherwise, directly write the data version and publication year.

Output

• Run records with queries, sources, timestamps, and failed items.
• A verified literature list marked with purposes, version relationships, and access levels.
• Conflicts, gaps, stop status, and next-round queries.

Stop Conditions

• Do not include items in the official literature list if basic bibliographic information cannot be verified.
• When systematic review-level completeness is required, do not only use open aggregation APIs.
• Do not bypass paywalls or access controls to obtain full texts.
• Do not treat search rankings, journal tiers, or author status as evidence levels.

Standalone Operation

• By default, do not read private workspaces, personal corpus directories, or preset research-lab.
• Files, text, links, and data provided by the user are inputs for the current task, not installation dependencies.
• When the network is unavailable, use stable methods within the Skill to continue; mark parts requiring current facts as pending verification.
• Can continue when other research Skills in this package exist; when installed alone, directly return instructions for the next step without stopping the current task.

Source Discipline

• Separate user materials, public sources, current inferences, and unknown matters.
• Reporting standards only check report transparency and do not automatically prove design quality.
• Abstracts from public sources are stored within the Skill; return to official pages when needing information on versions, fees, policies, APIs, or current journal status.
• Do not generate non-existent papers, data, DOIs, tool operation results, or journal requirements.