A research workflow should reduce repeated evidence handling and make provenance easier to inspect, not merely generate a plausible paper faster. This guide uses a fictional literature project on quiet areas in university libraries and students' learning experience. It separates search and citations, review synthesis, writing and Skill selection. Record identifiers are process examples; no specific studies, identifiers or effect estimates have been invented or verified here.
Identify the bottleneck and write a research task card
Choose the current problem: finding missing literature, resolving duplicate records, extracting evidence or organizing checked material into prose. State population, setting, phenomenon, dates, languages, databases and deliverable. The example might cover higher-education libraries, quiet-space use and self-reported learning experience rather than mixing all academic achievement, building acoustics and public-library studies without a rule.
Choose the review type and explain why. Rapid background research, scoping and systematic review have different aims and methodological demands. Tables and a flow diagram do not by themselves make a project systematic. Version the eligibility rules, evidence types, screening process and disagreement procedure before work. If evidence is limited, report that scope rather than widening the question until a strong conclusion becomes possible.
Keep raw exports read-only and separate screening, extraction and drafts. Record source and date for each file, with permitted use for unpublished notes, personal information and restricted full text. Inspect model services, retrieval interfaces, synchronization and logs separately. A local installation directory does not prove that evidence remains local. The team decides which material each service may process.
Stage one: retrieve literature and verify citation support
Turn the question into concept groups, synonyms, abbreviations and ambiguity exclusions. Log complete queries, fields, filters, dates, result counts and export files for each database. Syntax and disciplinary coverage differ; do not copy one query unchanged everywhere. Record interface errors, limits and unavailable full text rather than describing a failed retrieval as an absence of research.
Crossref, OpenAlex and PubMed have different roles and coverage, so select sources appropriate to the discipline. Metadata helps locate work; it does not establish relevance. An abstract cannot supply methodological detail that requires full text. Test whether a few known relevant works are retrievable and investigate misses. Record citation-chasing paths so a search does not remain confined to one citation network.
The record table preserves record_id, title, authors, year, source, DOI or other identifier, locator and access state. Normalize identifiers and titles for duplicate detection while mapping preprints, conference reports, journal versions and multiple reports of one study. Similar titles are not enough to delete a record. Distinguish records, reports and studies to avoid counting the same participants twice.
Citation verification has at least two checks: whether an identifier resolves to the correct work and whether the source context supports the manuscript claim. A working DOI, a high citation count or a well-formatted bibliography cannot establish support. Check authors, title, year and revision, including corrections or retraction information. Mark inaccessible full text as not checked rather than letting the agent guess page numbers and findings.
Zotero is an existing open-source reference-management option worth evaluating for records, attachments and citations rather than constructing a new bibliography system. Check version, license, synchronization and plugin access, particularly the audience of a shared group library. This article has not installed it, imported material or synchronized data, and does not present it as a verified SkillPI literature-search package.
Stage two: screen and extract before writing synthesis
Keep title-and-abstract and full-text screening records separately. Every decision needs the applicable rule, reason, reviewer and date. Send ambiguous cases to review rather than automatically excluding them. Calibrate reviewers on an overlapping sample and retain disagreements and adjudication. Unavailable full text differs from an ineligible study; flow counts should preserve that distinction.
An evidence matrix includes study and report identifiers, design, setting, sample, measurements, results, limitations, source locations and extractor. Preserve units, denominators, time points and uncertainty for key values. Not reported is not zero. Distinguish values read from figures, secondary summaries and authors' interpretations. A complete-looking table must not be achieved by model-imputed findings.
Use an appraisal approach suited to the research design and retain the supporting evidence for judgments rather than an unexplained star score. Surveys, interviews, experiments and observational work cannot be ranked only by participant count. Explain differences in setting, measures and design before deciding how synthesis is appropriate. A shared word such as experience does not justify pooled estimates or a claim of consistent effectiveness.
PRISMA 2020 provides reporting resources, including checklists and flow diagrams, not an accreditation badge or an automatically performed systematic search. Choose guidance appropriate to a scoping or other review where needed instead of treating one checklist as universal. Search, deduplication, screening, retrieval and inclusion counts must reconcile with the records; do not invent numbers to finish a diagram.
Worked example: link two reports of the same study
R-01 is a conference abstract and R-02 a later full article that may use the same participants. Compare authors, recruitment period, setting and sample description, record a suspected-same-study relationship and ask the researcher to adjudicate. If confirmed, retain both reports and their evidence roles but do not add their sample counts as independent studies.
If an abstract says experience improved but the full article reports a difference only in one subgroup, preserve subgroup, measure and uncertainty in the matrix. The synthesis must reflect that scope rather than citing the abstract to generalize to every student. Retain conflicts and resolution reasons instead of selecting the report that best supports the preferred argument.
A bounded agent request is: “Use approved-records and eligibility-plan-v1. Return record relationships, duplicate candidates, screening suggestions and an evidence matrix before prose. Give source locations for key facts and mark unreported fields. Do not create identifiers, full texts or statistical results. Do not contact authors, upload restricted material or submit a manuscript. Researchers decide final eligibility and interpretation.” Accept a small batch before extending the work.
Stage three: write from a claim matrix and preserve the handoff
Once evidence is stable, map each proposed claim to supporting studies, counterevidence, limitations and suitable wording. Synthesize by themes and methodological differences instead of stacking article summaries. Distinguish insufficient evidence from evidence of no effect and exploratory associations from causal findings. Outline, paragraphs and bibliography must use the same checked source set.
If the team already maintains approved reviewed notes in an authorized Notion workspace, Notion Research & Documentation fits this note-synthesis checkpoint. It can organize read pages around a question while preserving findings, caveats and source links. It needs working Notion tools, explicit readable sources and a writable destination. Without tools, a draft is not a saved page. It does not replace scholarly database retrieval, DOI checks, bias appraisal or final screening decisions.
Audit citation keys, in-text references and bibliography entries after drafting. Confirm that exported BibTeX or RIS files import into the destination tool. Follow the actual target style, including IEEE requirements when applicable, instead of judging by academic-sounding language. Style conversion does not verify source truth, and automated integrity checks do not remove responsibility for evidence and contribution.
Document contribution, methods, limitations and data or code availability, with AI-use disclosure following the institution and venue. The real team reviews authorship, design, data integrity, ethics and submission rules. Tools do not perform genuine peer review automatically or guarantee a submission-ready paper. Sending to a journal, emailing collaborators and changing sharing permissions remain separate actions; a finished draft does not imply that any occurred.
Discover and trial research Skills by stage
Search the SkillPI directory using discipline, source, stage, environment and output, such as education-record deduplication with DOI preservation and an auditable screening table. Compare actual source, version, license, dependencies and access before trying one nonsensitive reversible fixture. Download counts are not enough. Ensure that the discovered candidate and installation source refer to the same package rather than a same-name substitute.
For a discovery engine, verify the official SkillPI Skill Finder source and host compatibility, install through a supported host mechanism and confirm discovery before searching for the target capability. A finder identifies candidates; it does not prove safety, suitability or correct execution. The target package needs its own acceptance. Do not convert a competitor's catalog count, blanket security-review statement or eight source-article names into unverified SkillPI installation links.
Separate installation, host discovery, following the workflow and accepted artifacts. Use one small research question to check retrieval coverage, duplicate handling, source locators and file format. State database or full-text access gaps rather than inventing completion. As the catalog grows, rerun the fixture, add associations only at matching steps and record replacements and versions. The currently linked capability is verified note synthesis with its prerequisites, not an implied all-in-one scholarly engine.
FAQ and delivery checklist
Search or writing first? Stabilize retrieval, screening and evidence before drafting. Can an agent check every citation? It can assist with identifiers and matching, while source support, revision and use still need review. Does a PRISMA flow diagram make a systematic review? Transparent reporting is not proof that methods were correctly executed. Can unpublished drafts be supplied? Inspect services, logs, access and team authorization first, using approved summaries or an approved environment where necessary.
How should efficiency be assessed? Compare missed evidence, citation errors, rework and total time on comparable questions rather than generated word count. Deliver the plan, search log, raw exports, duplicate map, eligibility reasons, evidence and claim matrices, draft, bibliography and pending items. Keep unsupported claims in a verification queue so a researcher can locate the basis of every important conclusion and decide what follows.