open-scholar-skill

Draft, revise, or polish any section of a social science manuscript — Introduction, Theory, Methods, Results, Discussion, Abstract, or full paper. Saves draft sections to disk as publication-ready text and an internal writing log. Works best after /scholar-lit-review, /scholar-hypothesis, /scholar-design, and /scholar-analyze. Invoke with mode (draft/revise/polish), section name, topic, and target journal.

Design and execute computational social science analyses across 11 modules: text-as-data/NLP (STM, BERTopic, Wordfish, BERT, conText embedding regression, LLM annotation + DSL bias correction, XLM-R/mBERT); ML (supervised, Double ML, Causal Forests, Bayesian brms/Stan, conformal prediction); network analysis (ERGMs, SAOMs, relational event models, GNN via PyTorch Geometric); agent-based modeling (Mesa 3.x, NetLogo, LLM agents, ODD, SALib); computer vision (DINOv2, CLIP, ViT, multimodal LLMs, VideoMAE); LLM workflows (structured extraction, CoT coding, computational grounded theory, RAG); LLM synthetic data (personas, silicon sampling, vignette sim); geospatial (sf, tidycensus, spdep, Moran's I, spatial SEM); audio-as-data (Essentia, Whisper, pyannote, PANNs/wav2vec2); life-event sequence modeling (life2vec per Savcisens et al. 2024 NCS). Runs per-module verification subagents. Saves internal log + publication-ready Methods + Results. Targets NCS, Science Advances, and computational sociology venues.

Run data analytics and produce publication-quality tables and visualizations for social science research. Saves regression tables (HTML/TeX/docx), figures (PDF/PNG), an internal analysis log, and a publication-ready Results document (prose + table notes + figure captions). Accepts file paths, inline/pasted data, or fetches from online sources (NHANES, IPUMS, GSS, World Bank, etc.). Runs A9/B9 verification subagents to check analytic and visualization correctness. For causal designs, invokes /scholar-causal first. Use after /scholar-design.

Automatic quality auditor and continuous improvement engine for the scholar-skill suite. Four modes: (1) OBSERVE — post-skill output audit (run after any skill); (2) AUDIT — skill-suite structural health check; (3) IMPROVE — propose and apply fixes to skill definitions; (4) EVOLVE — cross-session pattern analysis and systemic improvements. Designed to run automatically at the end of any scholar-skill invocation.

Stable, deterministic social-science research-paper pipeline from idea or data to verified manuscript, citations, replication package, and final md/docx/tex/pdf outputs. Use when the user wants a reliable end-to-end auto-research chain that avoids the evolving scholar-full-paper workflow.

Generate research questions from existing materials — codebooks, survey questionnaires, datasets, or published papers/abstracts. Three modes: DATA (data files with safety scan + empirical signal tests), MATERIALS (codebook/questionnaire only with theory-driven ranking), PAPER (published paper/abstract → follow-up research ideas via SciThinker-30B + multi-agent evaluation). Auto-detects mode from file extensions. Explores the data landscape and proposes a ranked Top 10 list of publishable research questions using multi-agent evaluation. PAPER mode optionally uses SciThinker-30B (OpenMOSS-Team/SciThinker-30B on HuggingFace) for scientific ideation generation.

Comprehensive causal inference toolkit for social science research. Covers DAG construction, method selection, and thirteen identification strategies (OLS, DiD, RD, IV, FE, matching/reweighting, synthetic control, causal mediation, staggered DiD, DML/causal forests, bunching estimation, shift-share/Bartik instruments, distributional/quantile methods) — each with assumptions, diagnostics, R code, Stata code, and a write-up template. Use when the user needs to select and justify a causal identification strategy, build a causal DAG, or write the identification argument in their Methods section. Works between /scholar-hypothesis and /scholar-design.

Full citation management for social science manuscripts. Eight modes: (1) INSERT — add in-text citations to uncited claims; (2) AUDIT — find orphans, missing refs, year/author mismatches, duplicates; (3) CONVERT-STYLE — convert between styles (ASA, APA, numbered, author-date); (4) FULL-REBUILD — inventory claims → search Zotero + CrossRef → insert → assemble → audit; (5) VERIFY — verify each ref against local library (Zotero/Mendeley/BibTeX/EndNote), CrossRef, WebSearch; VERIFIED/UNVERIFIED/CORRECTED; (6) EXPORT — generate BibTeX .bib for LaTeX with cite keys, DOIs, enriched metadata; (7) RETRACTION-CHECK — cross-reference Retraction Watch; flag and suggest replacements; (8) REPORTING-SUMMARY — pre-filled NHB/NCS Reporting Summary (design, statistics, data availability) for Nature Human Behaviour and NCS. ABSOLUTE RULE: never fabricate — every reference verified against ≥1 authoritative database; local library searched first. Flags unsupported claims SOURCE NEEDED. Saves complete draft + audit log.

Systematic multi-agent code review of all analysis scripts produced in a project. 6 specialized agents review for: (1) correctness & logic errors, (2) robustness & defensive coding, (3) statistical implementation fidelity, (4) reproducibility & replication readiness, (5) code style & AI-generated anti-patterns, (6) data handling & variable construction (miscoded categories, wrong recodes, mishandled missing values, sample restrictions). Produces a consolidated review report with severity-ranked issues, fix checklist, and a per-script scorecard. Run after /scholar-analyze, /scholar-compute, or /scholar-eda to catch coding errors before manuscript drafting.

Multi-author research collaboration management. Covers CRediT role assignment, task delegation and tracking, co-author communication templates, contribution documentation, version management, conflict resolution, mentoring frameworks for student collaborators, and team coordination for multi-site or multi-PI projects. Use when managing a research team or preparing multi-author submissions.

Build original theoretical frameworks and produce publication-quality conceptual diagrams for social science research. Two modes: (1) THEORIZE — construct new theories from empirical puzzles using typology construction (property-space analysis), process theorizing, mechanism specification (Coleman's boat, Hedström DBO), scope condition mapping, multi-level models, abductive inference, and synthetic framework integration; (2) DIAGRAM — generate conceptual figures (mechanism diagrams, multi-level models, typology matrices, process/temporal models, concept maps, feedback loops, scope boundaries) as publication-ready TikZ/PDF or Mermaid/SVG with journal-specific formatting. Distinct from scholar-hypothesis (which selects FROM existing theories to derive testable hypotheses) — this skill builds the theories themselves. Distinct from scholar-causal (which builds DAGs for identification) — this skill represents conceptual relationships, not causal identification.

Comprehensive open data directory (100+ datasets across 14 categories) with auto-fetch capability, plus data collection instrument design, variable dictionaries, data management, IRB materials, and web/digital data collection for social science studies. Covers GSS, PSID, ACS, CPS, ESS, WVS, Afrobarometer, Eurobarometer, DHS, PISA, OECD, Eurostat, WHO, OpenAlex, Harvard Dataverse, ICPSR, Zenodo, OSF, and many more. Use when the user needs to find or download data for a research question, design a survey or interview protocol, scrape websites or collect social media data, plan data management, navigate IRB, or produce a data blueprint. Works best after /scholar-design and before /scholar-eda.

Plan a rigorous research design, run power analysis, specify variables, select an analytic strategy, design computational studies (NLP/ML/networks/ABM), write a pre-analysis plan, and draft the Data and Methods section for social science research. Produces a design blueprint saved to disk. Use after /scholar-hypothesis and before /scholar-data and /scholar-analyze.

Conduct exploratory data analysis (EDA) before hypothesis testing. Run when the user has a dataset and needs to load data, build the analytic sample, diagnose missing data, inspect distributions, visualize relationships, check collinearity, and produce a publication-ready Table 1 and pre-analysis memo. Produces figures (PDF/PNG) and tables (HTML/docx/TeX). Saves output to disk. Works after /scholar-data and before /scholar-analyze.

Research ethics toolkit for social scientists. Covers (1) AI tool data privacy audit — document and report how data was handled when using tools like Claude Code, Codex, or ChatGPT, and produce journal-required AI use disclosures; (2) plagiarism check — self-plagiarism, text recycling, AI-generated text detection guidance, and originality statements; (3) research authenticity audit — detect and remediate p-hacking, HARKing, data fabrication risks, selective reporting, and misinterpretation of results; (4) general ethics standards — IRB review, informed consent, CRediT authorship, conflict-of-interest disclosure, data sharing, and AI use declarations. Produces a saved ethics compliance report and declaration text ready to paste into submissions. Works at any stage; invoke before submission, after an ethics concern arises, or as part of the full-paper pipeline.

Develop theoretically grounded hypotheses and draft publication-ready Theory sections. Covers puzzle framing (5 types: anomaly, contradiction, extension, mechanism, scope), framework selection from 25+ sociological/linguistic theories (stratification, networks, culture, institutional, life course, analytical sociology via Coleman's boat + Hedström DBO, intersectionality, racial formation, social movements, status characteristics, signaling, non-Western: guanxi/Ubuntu/coloniality/world-systems), mechanism chain specification, hypothesis formalization (7 types: directional, interaction, mediation, curvilinear, comparative, boundary, null-as-finding) with formal and competing-prediction tables, text-based DAG construction (routes to /scholar-causal), and journal-calibrated drafts (ASR/AJS 1000–1500 words; Demography 600–1000; NHB/Science Advances 300–600 in Introduction; NCS 200–400; qualitative propositions). Saves internal theory log + section draft. Use after /scholar-lit-review and before /scholar-design.

Explore broad social science ideas and convert them into formal, researchable questions. Use when the user has an early-stage topic, puzzle, or rough question and wants sharper research questions, candidate mechanisms, scope conditions, testable hypotheses, and feasible study directions.

Initialize a new research project directory for open-scholar-skill. Creates the standard layout (data/raw, data/interim, data/processed, materials, output, .claude, logs), copies or symlinks raw files and materials into place, runs a local safety scan on every ingested file, and writes .claude/safety-status.json so the PreToolUse guard knows which files Claude may Read. Four modes: (1) init — create a fresh project from a slug and file list; (2) review — walk through NEEDS_REVIEW entries interactively and resolve each to CLEARED/LOCAL_MODE/ANONYMIZED/OVERRIDE/HALTED with logged rationale; (3) add — ingest new files into an existing project without rebuilding; (4) status — print the current state of .claude/safety-status.json and logs/init-report.md. Invoke before any data-touching scholar-* skill on a new dataset.

Format, select, and prepare a manuscript for journal submission. 5 modes: FULL-PACKAGE (structure audit + checklist + cover letter + open science), FORMAT-CHECK (audit against journal requirements), COVER-LETTER (journal-calibrated draft), SELECT-JOURNAL (score against 18 journals, ranked list + journal ladder), RESUBMIT-PACKAGE (post-rejection or R&R prep). Covers 18 journals: ASR, AJS, Demography, Du Bois Review, Science Advances, NHB, NCS, Social Forces, Language in Society, J. Sociolinguistics, Linguistic Inquiry, Gender & Society, APSR, JMF, PDR, SMR, Poetics, PNAS. Per-journal: word limits, section structure, abstract format, citation style, figure/table limits, open science (CRediT, COI, data/code availability, preregistration, Reporting Summary), blind review type, APC, submission URL, acceptance rate, turnaround. Includes cover letter templates, 8-dimension selection rubric, and open science package builder. Saves submission readiness report, cover letter draft, and open science declarations.

User-scoped, cross-project knowledge graph that persists extracted intellectual content (findings, mechanisms, theories, paper relationships) across projects and sessions. Layers on top of Zotero with richer content than raw bibliographic metadata. Eight modes: (1) INGEST — add papers from Zotero, PDF, lit-review output, DOI, or manual entry; (2) SEARCH — query by topic, author, theory, method, or finding; (3) RELATE — add or view inter-paper relationships (cites, contradicts, extends, replicates, uses-method, uses-theory); (4) STATUS — graph stats, recent additions, coverage; (5) EXPORT — project subset as markdown or NDJSON; (6) COMPILE — generate Obsidian-compatible markdown wiki (papers, concepts, topic clusters, contradictions, gaps); (7) ASK — answer research questions from the compiled wiki, save answers back; (8) RE-EXTRACT — re-run extraction on raw sources to upgrade papers or apply new schema. Storage: ~/.claude/scholar-knowledge/ (set via SCHOLAR_KNOWLEDGE_DIR). Raw sources archived in raw/.

Design and analyze studies in sociolinguistics, language variation, acoustic phonetics, discourse analysis, language contact, and computational linguistics. Covers VARBRUL/Rbrul for variation analysis, mixed-effects models for acoustic data, conversation analysis, critical discourse analysis, language attitudes (matched guise), corpus linguistics, computational sociolinguistics (conText embedding regression, LLM annotation for linguistic coding, BERT classification, semantic change detection, STM topic models), experimental sociolinguistics (factorial vignette experiments, IAT, reaction time paradigms, priming studies), Biber Multi-Dimensional Analysis (67 features, register comparison), and TTS-based matched guise tests. Produces R/Python code, publication-quality tables and figures, and saves output to disk. Use for Language in Society, Journal of Sociolinguistics, Language, Applied Linguistics, Nature Human Behaviour, Science Advances, and Nature Computational Science.

Conduct a systematic literature review AND develop theory and hypotheses in one integrated workflow. Use when the user wants to map existing knowledge, identify gaps, select theoretical frameworks that address those gaps, specify mechanisms, and derive testable hypotheses — producing a publication-ready Literature Review and Theory section. Works best after /scholar-idea. Replaces running /scholar-lit-review and /scholar-hypothesis separately.

Conduct a systematic literature review that maps the current landscape of a research question. Produces a structured literature landscape map (field evolution, theoretical landscape, established/contested/null findings, mechanism inventory, methodological landscape, gap analysis), a publication-ready review draft, and a search log. Three modes: full landscape map (40-80 papers), targeted review for a paper's intro (15-30 papers), or rapid scoping review (10-20 papers). Searches local reference library (Zotero/Mendeley/BibTeX/EndNote) first, then multi-wave web search, Annual Reviews checkpoint, and citation chain expansion. Runs a verification subagent and saves two output files to disk. Use when the user wants to map the existing literature, identify research gaps, synthesize findings, or build a foundation for a paper. Works best before /scholar-hypothesis or /scholar-lit-review-hypothesis.

Current-awareness literature monitoring for social sciences. Tracks new publications from top journals (ASR, AJS, Demography, Social Forces, NHB, NCS, Science Advances, APSR, PNAS), preprint servers (arXiv cs.CL for LLM research, cs.CY for computational social science, econ.GN), and custom author/keyword watchlists. Delta-based and /loop-safe: each invocation fetches only papers published since the last tick, groups them by user-defined category, generates 2–3 sentence LLM summaries of title+abstract, and auto-ingests into the scholar-knowledge graph. Delivers digests to the user's phone via Telegram (using the existing MCP channel) or ntfy.sh push, optionally email over SMTP — with a file-based audit trail always written to output/monitor/. Modes: default fetch (all due enabled sources), targeted fetch (source_id argument), all (force-fetch all enabled), preview (dry-run — show what would fetch without network calls), init, list, status, add, remove, configure delivery, digest. Designed for /loop (e.g., '/lo

Implement open science practices for a social science study. Five modes: (1) PREREGISTER — OSF / AsPredicted / EGAP / Registered Reports; secondary-data preregistration; deviation reporting; (2) DATA-SHARE — FAIR principles, de-identification, repository selection (Dataverse / ICPSR / Zenodo / QDR), social media policies, computational data (model weights, embeddings), data availability statements; (3) CODE-SHARE — minimum + gold-standard replication packages, README template, renv / conda / Docker / Makefile environments, CITATION.cff, Zenodo DOI via GitHub; (4) FULL-PACKAGE — data management plan (NSF/NIH template), CRediT authorship, COI, IRB statement, preprint strategy, open access + APC waiver guidance; (5) REPLICATION-PACKAGE — end-to-end audit against journal requirements. Required at Nature journals; increasingly expected at ASR, Demography, and Science Advances. Saves preregistration document, data management plan, and replication-package README to disk.

External review via OpenAI Codex CLI agents. Spawns multiple parallel Codex agents (codex exec) to independently review analysis scripts, verify manuscript-to-output consistency, check statistical logic, and audit reproducibility. Each agent writes a review report to disk. Claude reads all reports, synthesizes a consolidated review with severity-ranked issues, and presents a fix checklist. Read-only — diagnoses but does not modify any project files.

Personalize and refine academic manuscript writing style by injecting distinctive authorial micro-patterns (tiny topos). Detects and replaces generic hedging stacks, formulaic transitions, symmetrical parallelism, over-enumeration, and other flat prose patterns with natural, voice-rich academic writing. Three modes: SCAN (diagnose style patterns), REWRITE (apply fixes), FULL (scan + rewrite). Preserves argument structure, citations, and technical content while elevating prose to a distinctive, polished authorial voice.

Qualitative research methods toolkit for social sciences. Covers codebook development, grounded theory coding (open, axial, selective), Braun & Clarke reflexive thematic analysis, systematic content analysis (Krippendorff), LLM-assisted qualitative coding with human validation (Lin & Zhang 2025 framework), mixed-methods integration, and inter-coder reliability assessment. Produces codebooks, coded datasets, code-to-quote mappings, thematic maps, reliability reports, and publication-ready qualitative write-ups. Exports to NVivo, Atlas.ti, Dedoose, and MAXQDA formats.

Build, document, test, and validate a journal-ready replication package for social science research. Six modes: (1) BUILD — assemble scripts, data, documentation into structured replication directory from output/[slug]/scripts/ and project artifacts; (2) DOCUMENT — generate comprehensive README (AEA template), codebook, data dictionary, computational requirements, dependency graph; (3) TEST — execute clean-run validation in isolated environment (fresh renv/conda, Docker optional), compare reproduced outputs against published tables/figures; (4) VERIFY — paper-to-code correspondence audit ensuring every table, figure, and in-text statistic has a producing script; (5) ARCHIVE — create versioned release, Zenodo DOI, GitHub integration, repository deposit preparation; (6) FULL — run all modes sequentially. Consumes output/[slug]/scripts/ from scholar-analyze/scholar-compute. Targets ASR, AJS, Demography, Science Advances, NHB, NCS, APSR.

Simulate peer review, draft point-by-point responses to reviewer comments, revise a manuscript, plan a resubmission to a new journal after rejection, or write an R&R cover letter. 5 modes — simulate (3–4 parallel journal-calibrated reviewer agents + severity matrix + revision roadmap), respond (categorized triage dashboard + point-by-point letter + changes summary table), revise (word-budget-tracked section edits via /scholar-write), resubmit (rejection diagnosis + journal retargeting + cover letter), cover-letter (standalone R&R or resubmission cover letter). Supports multi-round R&R tracking. Saves response letter, revision plan, and cover letter to disk.

Real-time data privacy and leakage protection layer for AI-assisted research. Before any file is read by Claude Code and transmitted to Anthropic's API, proactively scans it locally for sensitive content (PII, HIPAA, IRB-protected, restricted licensed data) using local Bash pattern matching — so the sensitive data itself never enters the AI context during the scan. Issues tiered warnings (green/yellow/red), requests explicit user permission before transmitting any sensitive data, and offers safe alternatives: anonymization scripts, local-mode analysis (bash-only, no data in context), data minimization, or full halt. Four modes: SCAN (audit a file using only local grep/awk — sensitive data stays local), GATE (intercept a proposed operation and ask permission before execution), PROTOCOL (generate a project-level data safety protocol), STATUS (safety status log). Works alongside scholar-eda, scholar-analyze, scholar-compute, and scholar-data to protect any data-touching operation.

Run LLM-powered social simulations at scale: silicon sampling, generative ABM, survey/vignette/conjoint experiments, opinion dynamics. Multi-provider (Anthropic/OpenAI/open-source/local). Validation against human data is mandatory before any publishable claim.

Two-stage verification of analysis-to-manuscript consistency using a 4-agent panel. Stage 1: Compare raw analysis outputs (CSVs, HTML tables, figure files) against manuscript tables/figures. Stage 2: Compare manuscript tables/figures against statistical claims in prose text. Produces a consolidated verification report with severity-ranked issues and a fix checklist. Run after scholar-write or before scholar-journal submission prep.

Synchronize content across presentation slides, speaker script, and manuscript/paper. Audits for stale references, numbers, citations, and version mismatches, then updates all files in parallel and compiles to PDF.