Imported from Sheldon-92/TAD (
.agents/skills/academic-research/SKILL.md). Install upstream withnpx skills add Sheldon-92/TAD --skill academic-research. Copyright stays with the author.
Academic Research Capability Pack
Teaches AI agents HOW to do academic research — depth enforcement, citation integrity, quality scoring, and self-evaluation. Derived from ScienceClaw's 629-line SCIENCE.md protocol + ScholarEval rubric + VOYAGER-inspired Reflexion Cycle.
CONSUMES: Research question + optional domain constraints (field, date range, databases). PRODUCES: Evidence-grounded research report + methodology section + citation audit trail + Reflexion Cycle self-evaluation (for non-trivial tasks).
Scope Disambiguation
This pack covers academic and scientific research methodology.
This pack — use when:
- User asks about 学术/academic/论文/paper/文献/literature tasks
- Systematic or literature reviews, meta-analyses
- Tasks mentioning PRISMA, PubMed, citation analysis, ScholarEval
- Evaluating research quality or scholarly writing
NOT this pack — defer to *research when:
- Market analysis, technology trend analysis, competitive analysis
- General non-academic "research X" tasks without scientific context
- Business strategy investigation, product analysis
If ambiguous, ask the user: "Is this academic/scientific research or general market/technology research?"
Step 1: Detect Research Task Type
Classify the user's request into one of four tiers. Each tier has a mandatory minimum depth derived from ScienceClaw SCIENCE.md lines 111-121.
| Tier | Signal Keywords | Min Phases | Min Tool Calls | Session Model |
|---|---|---|---|---|
| Quick factual | "what is", "define", "quick question", single-fact lookup | 1-2 | 3-5 | Single session, no handoff |
| Literature survey | "review literature", "survey papers", "what does the research say", "文献综述" | 1-5 | 20-40 | 1-2 sessions with handoff |
| Comprehensive review | "comprehensive review", "analyze the field", "综合分析", "学术综合分析" | 1-6 | 40-80 | Epic: 2-3 phases |
| Systematic review | "systematic review", "meta-analysis", "PRISMA", "系统性综述", "元分析" | 1-6 (iterated) | 80+ | Epic: 4-6 phases (PRISMA pipeline) |
| Multimodal research | user uploads image, "pattern analysis", "visual comparison", "artifact image", "图像分析", "纹样" | 1-5 | 20-40 | 1-2 sessions; load multimodal-research.md + pattern-extraction.md |
Example inputs per tier:
Quick factual:
- "What is the current impact factor of Nature Medicine?"
- "Who first described CRISPR-Cas9 gene editing?"
- "What is the sample size formula for a two-arm RCT?"
Literature survey:
- "Survey recent papers on transformer architectures for protein folding"
- "What does the literature say about remote work productivity?"
- "Review papers on CRISPR cancer therapy from 2023-2025"
Comprehensive review:
- "Analyze the field of large language models for drug discovery"
- "Comprehensive review of causal inference methods in economics"
- "Review all approaches to federated learning in healthcare"
Systematic review:
- "Conduct a systematic review of CBT interventions for anxiety"
- "PRISMA-compliant meta-analysis of statin efficacy in primary prevention"
- "Systematic review of AI diagnostic accuracy in radiology"
Multimodal research:
- "Analyze this image of a ceramic bowl — describe the ornamental pattern and classify the motif types"
- "Compare the palmette designs across these three artifacts from different cultures"
- "Extract the visual features from this food plating photograph for sensory evaluation research"
Source: Adapted from SCIENCE.md lines 111-121, adjusted per tad-mapping-blueprint.md Decision 6
Step 2: Load References by Task Type
Protocol References (always applicable)
| Tier | Protocol References |
|---|---|
| Quick factual | zero-hallucination.md only |
| Literature survey | research-protocol.md + zero-hallucination.md + fallback-chains.md |
| Comprehensive review | ALL protocol references + relevant cluster references |
| Systematic review | ALL protocol + ALL cluster references |
Cluster References (load by research domain)
Load the cluster references that match the research topic. Multiple clusters may apply.
| Research Domain Signal | Load These Cluster References |
|---|---|
| Literature search, citation analysis | references/literature-search.md |
| Academic database queries (general) | references/database-apis-general.md |
| Life science databases (protein, gene, drug) | references/database-apis-life-sciences.md |
| Statistical analysis, meta-analysis | references/statistics.md |
| Paper/grant/report writing | references/writing.md |
| Figures, plots, charts | references/visualization.md |
| Biomedical / life science domain | references/domain-biomedical.md |
| Physical / computational science | references/domain-physical.md |
| Social science / economics | references/domain-social.md |
| Experiment design, ethics, reproducibility | references/experiment-design.md |
| Image analysis, visual evidence, multimodal | references/multimodal-research.md |
| Ornamental patterns, motifs, cross-cultural visual comparison | references/pattern-extraction.md |
| Quantitative image measurement, CV tools, similarity scoring, frequency analysis | references/quantitative-analysis.md |
Read the loaded reference files and apply their rules during research execution.
Step 3: Alex/Blake Role Mapping for Research Tasks
| Research Activity | Alex (Design) | Blake (Execute) |
|---|---|---|
| Research question formulation | Elicits via Socratic Inquiry | — |
| Search strategy + database selection | Specifies in handoff | — |
| Search execution | — | Runs curl/API commands |
| Paper reading + extraction | — | Reads and extracts findings |
| Quality assessment | Defines criteria in handoff ACs | Applies ScholarEval from scholar-eval.md |
| Synthesis across sources | — | Synthesizes with citation audit |
| Report writing | — | Writes report with methodology section |
| Depth enforcement | Sets minimum tier in handoff | Follows research-protocol.md thresholds |
| Citation integrity | AC: "every citation traces to tool result" | Self-checks per zero-hallucination.md |
Source: tad-mapping-blueprint.md Decision 5
Step 4: Anti-Premature-Conclusion Checklist
Before concluding ANY research task (except quick factual), apply the 10 Anti-Premature-Conclusion Rules from research-protocol.md and the Completeness Checklist from the same file.
If any item is unchecked, continue working instead of concluding.
Anti-Skip / Anti-Rationalization Table
The agent WILL be tempted to stop early. Each rationalization below is a known failure mode; the counter cites the existing protocol that forbids it. Do NOT proceed past a row whose left column matches your current thought.
| Rationalization (the excuse) | Counter (the rule it violates) |
|---|---|
| "The first 3 papers agree, I can stop." | Premature saturation. research-protocol.md Rule 1 (≥2 databases) + Rule 9 (count tool calls vs the tier floor in Step 1: 20-40 survey / 80+ systematic). 3 concordant papers ≠ exhaustive search. |
| "Tier is Quick factual, so I can skip ScholarEval / the depth floor." | Re-classify, don't downgrade. If the request says "review / survey / systematic / meta-analysis" it is NOT Quick factual — match Step 1 signal keywords honestly. ScholarEval ≥ 0.75 is the Gate 3 Accept bar for any non-trivial tier. |
| "The API returned empty, so the topic must have no literature." | Empty result ≠ no literature. fallback-chains.md 3-strike rule: try the alternative database/query before declaring a source failed. zero-hallucination.md FORBIDS substituting training knowledge for a 0-result search. |
| "The abstract said so, I'll cite it from memory." | zero-hallucination.md 4-point check: every citation detail (title/authors/DOI/count) MUST come from a tool result in THIS session. Reciting from memory = fabrication. |
| "Close enough — I'll write the report now without a Methods section." | research-protocol.md Rule 4: no report without a Methods section (databases queried, result counts, filter criteria). Rule 8: do not end the turn until the report is saved to a file. |
| "Cohen's kappa is fine; I don't need the other agreement metrics." | For systematic-review screening the include class is rare → kappa paradox. literature-search.md §2 requires Gwet's AC1 + raw % agreement alongside kappa when the positive class is rare. |
| "PRISMA is PRISMA — the 27-item checklist covers this scoping review." | Wrong instrument. literature-search.md PRISMA decision rule: scoping review → PRISMA-ScR (20+2 items); only a focused-question systematic review uses PRISMA 2020 (27 items). |
These rows are referenced from the Quick Rule Index below. The full 10 Anti-Premature-Conclusion Rules live in research-protocol.md.
Step 5: TAD Integration
| TAD Mechanism | Academic Research Role |
|---|---|
| Gate 3 | Verifies citation integrity (zero-hallucination 4-point check) + ScholarEval score ≥ 0.75 for Accept |
| Gate 4 | Verifies research completeness per tier (did Blake reach the minimum phase and tool-call threshold?) |
| Knowledge Assessment | Runs after Reflexion Cycle for non-trivial tasks — captures reusable research patterns in project-knowledge |
| Ralph Loop Layer 1 | Checks fallback chain exhaustion (did Blake try alternatives before declaring a source failed?) |
Step 6: Research Memory & Persistence
Research findings persist across sessions via TAD's existing memory stack — no additional infrastructure required.
| Memory Need | TAD Solution | How It Works |
|---|---|---|
| Cross-session findings | .tad/project-knowledge/*.md |
Knowledge Assessment (Gate 3/4) writes reusable discoveries to category files. Blake or Alex reads them on next session via @import in CLAUDE.md. |
| Semantic recall across sources | NotebookLM via *research-notebook ask |
Multi-source notebooks (30-50 sources each) enable cross-source synthesis. Query returns grounded answers with source attribution. |
| Per-topic evidence archive | .tad/evidence/research/{topic}/ |
Subdirectories per research topic store raw findings, analysis files, and search logs. Persists across handoffs within an Epic. |
| Self-evaluation patterns | Completion report Knowledge Assessment | Reflexion Cycle (reflexion-cycle.md) captures what worked and what failed. Written to completion report, promoted to project-knowledge if reusable. |
| Research notebook portfolio | .tad/research-notebooks/REGISTRY.yaml |
Index of all NotebookLM notebooks with topic, source count, and status. Alex creates and curates; Blake queries during implementation. |
Integration rules:
- Findings that apply beyond a single handoff → write to
.tad/project-knowledge/via Knowledge Assessment - Findings specific to one research question → keep in
.tad/evidence/research/{topic}/ - Findings worth cross-source synthesis → ingest into NotebookLM via
*research-notebook ingest - Do NOT duplicate: choose one persistence path per finding
Quick Rule Index
Before concluding, re-read the Anti-Skip / Anti-Rationalization Table in Step 4 — it maps each "stop early" excuse to the rule that forbids it.
Protocol References (Phase 2)
| Reference | What It Covers | When to Read |
|---|---|---|
| research-protocol.md | 6 mandatory phases, depth enforcement, mandatory search protocol | Every non-trivial research task |
| zero-hallucination.md | Citation integrity, 4-point self-check, empty-result handling | Every task — this is absolute |
| scholar-eval.md | 8-dimension weighted quality rubric (0-1 scale) | When evaluating research quality |
| reflexion-cycle.md | 5-dimension post-task self-evaluation | After completing a research task |
| fallback-chains.md | Source failure recovery, 3-strike rule, fallback tables | When a search/API fails |
Cluster References (Phase 3) — Domain-Specific Judgment Rules
| Reference | What It Covers | When to Read |
|---|---|---|
| literature-search.md | Multi-database search, PRISMA pipeline, citation networks, bibliography management | Literature reviews, systematic reviews |
| database-apis-general.md | Semantic Scholar, OpenAlex, PubMed, arXiv, World Bank, CrossRef API templates | Any database query task |
| database-apis-life-sciences.md | UniProt, ChEMBL, NCBI, PDB, ClinicalTrials, KEGG, STRING API templates | Life science research |
| statistics.md | Test selection, meta-analysis (DerSimonian-Laird, I²), effect sizes, power analysis, APA reporting | Any quantitative analysis |
| writing.md | IMRaD structure, grant writing (NIH/NSF), LaTeX, citation styles, journal page limits | Writing research outputs |
| visualization.md | Publication figures (300+ DPI), journal palettes, chart selection, statistical plots | Creating figures |
| domain-biomedical.md | Bioinformatics (FDR<0.05), clinical trials, CONSORT, drug discovery, protein analysis | Biomedical research |
| domain-physical.md | Molecular dynamics, materials screening, signal processing, computational chemistry | Physical/computational science |
| domain-social.md | Econometrics (DiD, RDD, IV), survey methods, psychometrics, education research | Social science research |
| experiment-design.md | RCT design, sample size, GRADE, Cochrane RoB, reproducibility, IRB, peer review | Experiment planning |
| multimodal-research.md | Image analysis protocol, observation-before-interpretation, measurement fallbacks, image citation, cross-image comparison | Image-based research tasks |
| pattern-extraction.md | Motif identification, line abstraction (3 levels), cross-cultural feature matrix, ornamental vocabulary (guilloche, arabesque, etc.) | Visual pattern / ornament studies |
| quantitative-analysis.md | image-analysis.py decision matrix, similarity/frequency thresholds, output interpretation, qualitative-quantitative integration | When using CV tools for measurement |
Available Tools
scripts/academic-search.sh — Database Query Helper
Reusable bash script for querying academic databases with rate limiting and structured output.
academic-search.sh <database> "<query>" [--limit N]Supported databases: semantic-scholar, openalex, pubmed, arxiv, europeana (requires EUROPEANA_API_KEY), usda-food (DEMO_KEY or USDA_API_KEY).
scripts/image-analysis.py — Quantitative Image Analysis
Python toolkit for CV-based image measurement. Requires setup: bash scripts/setup-cv.sh
~/.academic-research-cv-venv/bin/python3 scripts/image-analysis.py edges input.jpg --output edges.svg
~/.academic-research-cv-venv/bin/python3 scripts/image-analysis.py colors input.jpg --output colors.json
~/.academic-research-cv-venv/bin/python3 scripts/image-analysis.py match img1.jpg img2.jpg --output result.json
~/.academic-research-cv-venv/bin/python3 scripts/image-analysis.py frequency input.jpg --output freq.json
~/.academic-research-cv-venv/bin/python3 scripts/image-analysis.py features input.jpg --output features.jsonSee quantitative-analysis.md for decision matrix and output interpretation.
Use this script during research Phase 1 (Discovery) and Phase 4 (Database Cross-Verification) instead of writing raw curl commands.
Notes
- Skill evolution: This pack improves via TAD's existing *optimize → proposal → human approval → handoff cycle (not runtime generation). See tad-mapping-blueprint.md Decision 4
- Memory: Uses TAD's file-based project-knowledge + optional NotebookLM notebooks. See Step 6 for the full memory integration mapping
- Source coverage: 86 unique ScienceClaw skills cited across 15 reference files (consolidated from 150 P1+P2 skills)