Academic Skills for Food & Nutrition Science

AI research assistant for food, nutrition and agricultural science — Claude Code, Codex, MiniMax Agent, and OpenClaw skills for literature review, systematic review (PRISMA & meta-analysis), data analysis and statistics, scientific figures, journal formatting, and peer review. Food, nutrition & agriculture research automation, end to end.

Original, MIT-licensed skills for the food, nutrition & agriculture research lifecycle — research → write → review → revise → finalize — where each core skill is a multi-subagent system and a master pipeline orchestrates them, with built-in knowledge of food, nutrition & agriculture journal author guidelines and a scientific figure workflow. Two parallel skill sets share one engine: food-* for food & nutrition science and agri-* for agricultural science. Supports Claude Code, Codex, MiniMax Agent, and OpenClaw.

This open project was initiated by the Food Science Group at the University of Melbourne, and we warmly welcome food & nutrition research groups from around the world to use, adapt, and contribute to it. MIT-licensed and open source.

Install

Claude Code (one command):

claude plugin marketplace add PangenomeAI/academic-skills-food-nutrition && \
claude plugin install academic-skills-food-nutrition@academic-skills-food-nutrition

Then restart Claude Code (or run /plugin). Update later with claude plugin update academic-skills-food-nutrition.

Claude Code, Codex, and MiniMax Agent (one command via the installer):

curl -fsSL https://raw.githubusercontent.com/PangenomeAI/academic-skills-food-nutrition/main/install.sh | bash

Or, from a local clone: ./install.sh (all) · ./install.sh claude · ./install.sh codex · ./install.sh minimax · ./install.sh openclaw. The installer registers the Claude Code plugin, and for Codex (${CODEX_HOME:-~/.codex}/skills/), MiniMax Agent (Mavis; ${MAVIS_SKILLS_DIR:-~/.mavis/skills}/), OpenClaw (${OPENCLAW_HOME:-~/.openclaw}/skills/) it installs each skill flat (…/skills/<name>/SKILL.md, so the agent discovers it) plus the shared journals/ and scripts/ directories so cross-skill references resolve. Restart the app so it rescans skills. (Override the location with CODEX_HOME / MAVIS_SKILLS_DIR / OPENCLAW_HOME, or add via MiniMax's in-app Skill Creator/import.)

Updating

Updates are not automatic — once installed you stay on that version until you choose to update. To get the latest release:

Claude Code (built-in update):

claude plugin update academic-skills-food-nutrition

Restart Claude Code (or run /plugin) to apply. If it doesn't yet show the newest version, refresh the marketplace first:

claude plugin marketplace update academic-skills-food-nutrition

Codex / MiniMax Agent (Mavis) / OpenClaw — re-run the installer (it cleanly replaces the skills), then restart the app:

curl -fsSL https://raw.githubusercontent.com/PangenomeAI/academic-skills-food-nutrition/main/install.sh | bash
# or, from a local clone:  ./install.sh codex  |  ./install.sh minimax  |  ./install.sh openclaw

Check your installed version with claude plugin list; see all versions on the Releases page (a GitHub Release is published automatically on each feature update).

Skills

Core workflow

• food-research — comprehensive, multi-source literature discovery and evidence synthesis for food & nutrition (FSTA, PubMed, Web of Science, Scopus, AGRICOLA, preprints, semantic search; EFSA/FDA/USDA/Codex for safety and regulatory evidence). Four-layer search, two-phase screening, and synthesis via subagents; grades evidence and maps gaps. Four streams — quick brief, full review, deep research, systematic. The first three prioritize sources by journal ranking (journal_ranker: Q1/Q2 food-science & nutrition, plus Nature/Science/Cell families and Q1/Q2 in any other discipline = highest priority; Q3 second; Q4 avoided). The full review and systematic streams finish by writing a manuscript, running an editorial + integrity reviewer loop, and exporting a Word (.docx) (APA 7.0 default, or a target journal via journal-selector). The PRISMA 2020 systematic review stream adds a fixed protocol, ≥3 databases (Web of Science/Scopus/PubMed), dual independent three-step screening (title → abstract → full text) with a moderator, a PRISMA flow diagram, a results table, and OHAT risk-of-bias (in vitro / human / animal); it uses eligibility-based inclusion rather than journal ranking.

• food-deep-research — source-validated literature-review engine (scope → design → discover → screen by journal ranking → validate every source → extract & verify evidence → synthesize → stress-test → write & format → editorial + integrity review loop) with a 12-subagent team. Outputs a finished, formatted review (APA 7.0 by default, or a target journal's style via journal-selector). Runs standalone or as food-research's deep-dive engine.

• food-paper — whole-process manuscript system (12 subagents) covering the full research lifecycle: understand the field (calls food-research), frame research questions, curate data, run statistics, build figures & tables (calls food-figure), construct arguments and discussion, draft, polish, manage citations, and self-review (calls food-review) — journal-aware throughout (APA 7.0 default, or a target journal via journal-selector).

  8 modes — you rarely need the whole pipeline; the two most-used are polish and format-convert. See food-paper modes.

• food-review — multi-reviewer peer-review panel (coordinating editor + methodology, domain/novelty, and integrity/ethics reviewers + a devil's advocate) with a formatting-compliance check against the target journal (APA 7.0 default, or a specific journal via journal-selector), ending in an editorial decision + revision checklist + response-letter skeleton.

• food-pipeline — master orchestrator that routes a project to the specialist skills (each with its own subagent team) and enforces quality gates: journal selection → research (food-research/food-deep-research) → write & analyze (food-paper → food-figure) → peer review (food-review) → revise → finalize, with mandatory author decision points. Default: one review→revise round; a second round and in-place Tracked Changes on the original Word file require explicit author authorization (avoids over-automation).

food-paper modes: what each one does

food-paper is eight tools in one, and you rarely need the whole pipeline. You never type a mode name: ask in plain English and the right mode is selected. agri-paper has the same eight modes for agricultural manuscripts.

⭐ = the two most-used. Both work on a manuscript you already have, and neither rewrites your science.

What polish actually does

Beyond grammar and clarity, its polisher subagent runs human-writing.md to strip the tells that make text read as machine-generated:

• inflated significance ("plays a pivotal role", "paves the way")
• vague attribution ("studies have shown" → a real citation)
• stock AI vocabulary ("delve", "intricate landscape", "showcase")
• copula avoidance ("serves as" → "is"), filler, stacked hedging
• generic upbeat endings ("opens exciting avenues")

…then asks "what still reads as machine-written?" and fixes what it finds.

It deliberately keeps calibrated hedging (deleting a "may" your data require would be a fabrication-grade error, not a polish), passive voice in Methods, and the journal's required heading style — and it never changes a number, a claim's scope, or a citation. This is a writing-quality edit, not a way to hide AI use: the AI-use disclosure requirement stands regardless of how the prose reads.

Agricultural science (agri-*)

The same five-skill workflow for agricultural science, acting as a senior agricultural scientist of the relevant discipline — agronomy · soil science · horticulture · dairy & animal science · agricultural engineering · agricultural economics & policy · agriculture multidisciplinary.

• agri-research · agri-deep-research · agri-paper · agri-review · agri-pipeline — each delegates its machinery to the corresponding food-* skill (same subagents, gates, modes, and output contracts) and applies three substitutions: the persona, the evidence base, and journal routing. One engine, no duplicated architecture. Figures still go through food-figure, which is domain-neutral.
• Evidence base: agriculture + multidisciplinary literature, ranked like the food suite — Tier 1 the 230 Q1/Q2 agriculture journals plus the Nature/Science/Cell/PNAS families and Q1/Q2 adjacent disciplines; Tier 2 Q3 for gaps only; Q4 avoided. FAO, USDA, CGIAR, EFSA and extension sources count as evidence with a source and date.
• Agricultural rigour: field-trial reporting (site, season/years, soil, cultivar, design, replication), the experimental unit (pseudoreplication is the classic fatal flaw), G×E and season-to-season variation, ARRIVE for animal work, soil sampling depth and equivalent-soil-mass basis, and no pot-to-field extrapolation. The contract: agri-research/references/agriculture-domain.md.

Journal knowledge

• journal-selector (shared machinery — deliberately not in the skill list) — asks which journal you're targeting (or reads it from your request) and loads that journal's constraints. You never invoke it directly: just name a journal to any skill above and it runs. Covers the Food Science & Technology (60) and Nutrition & Dietetics (59) journal lists, the 230 Q1/Q2 agriculture journals across seven JCR categories, plus 35 multidisciplinary / cross-discipline journals these researchers publish in (Nature, Science, Cell, and PNAS families, eLife, PLOS, ES&T, Gut, etc.). See journals/_coverage.md, journals/_coverage_nutrition.md, journals/_coverage_agriculture.md, and journals/_coverage_multidisciplinary.md.
• journals/* — 25 publisher-tiered author-guideline skills covering the Food Science & Technology, Nutrition & Dietetics, agriculture (Q1/Q2), and multidisciplinary journal lists (Elsevier, Wiley, Nature Portfolio, Springer, Taylor & Francis, MDPI, RSC, ACS, Annual Reviews, Oxford, Emerald, KeAi/Tsinghua, Codon, BioMed Central, Cambridge, Frontiers, plus niche-publisher skills for nutrition and agriculture). Because the folders are publisher-tiered, an Elsevier agronomy journal reuses the same Guide-for-Authors skill as an Elsevier food journal — 185 of the 230 agriculture journals need no new format. Each lists the journals it covers, their limits, structure, reference/citation style, and a submission checklist.

Figures

• food-figure — comprehensive figure system: analyzes your data (scripts/analyze_data.py profiles a CSV/TSV and recommends the best figure type), then renders submission-grade graphics in Python or R at the target journal's spec. Covers all common scientific figure types (bar/box/violin, line/kinetic, scatter/regression, Bland-Altman, sensory radar, chromatograms, TPA/rheology, dose-response, survival, PCA/PLS-DA, clustered heatmaps, forest, microscopy plates, multi-panel), with Python (matplotlib/seaborn/subplot_mosaic/ statsmodels) and R (ggplot2/patchwork/ComplexHeatmap/ggrepel + svglite/cairo_pdf/ ragg) template libraries, curated colourblind-safe palettes, per-figure provenance + captions, and a QA gate. Exports journal-ready SVG/PDF/TIFF. For schematics/graphical abstracts, an opt-in AI-image route (Gemini/ChatGPT/ Claude) with a structured prompt method — never for data figures. A runnable synthetic/illustrative Figure-story gallery provides four final 11- or 12-panel food-science figures spanning active packaging, probiotic storage, trained sensory analysis, and analytical-method validation. Each narrative progresses from a code-drawn experimental schematic through topic-specific raw and derived evidence to an explicitly illustrative synthesis panel, with deterministic source-data generation, PDF + PNG exports, captions, and trace cards. The gallery also documents an AI schematic-route test and its QA/reproducibility safeguards.

How it fits together

Name a journal ("I want to publish on Food Chemistry" / "format for LWT") and journal-selector loads that journal's rules; food-paper writes to them and re-flows the reference list into the journal's citation style; any figure request goes to food-figure at the journal's DPI and column width. Ask for the whole thing and food-pipeline runs research → write → review → revise with checkpoints.

Which one do I pick?

Use the smallest skill that does the job — the pipeline is for whole projects, not one-off tasks.

Reformatting is deliberately not a pipeline job: running the pipeline on a finished paper would re-research, re-review, and re-edit work you already consider final.

Coverage

Every target journal maps to a publisher-tiered skill, verified by scripts/check_journal_coverage.py:

The agriculture list is the Q1 and Q2 journals of all seven JCR agriculture categories (Agronomy · Agriculture Multidisciplinary · Dairy & Animal Science · Soil Science · Agricultural Economics & Policy · Horticulture · Agricultural Engineering), from JCR 2025, deduplicated across categories at each journal's best quartile. Q3 fills gaps only and Q4 is avoided. Journals already in the food or nutrition lists keep their existing folder (Journal of Dairy Science → j-dairy-science), and 185 of the 230 reuse an existing publisher format — only society and regional titles fall to agri-other.

Author-guideline details record a Source: URL and a Verified: date. Publisher pages change and several block automated access — confirm exact numeric limits at the source before submitting; structure and reference styles are the stable part.

Using AI responsibly for academic work

AI can make mistakes. Large language models can produce fluent but incorrect statements, invented references, and unsupported conclusions. You — the researcher — are responsible for the integrity of your work.

These skills are research assistants, not authors. Treat every AI-generated statement as a draft to be verified, never as established fact.

Your responsibilities

• Validate everything. Check every claim, number, and citation against the primary source before you use or submit it. This project enforces an anti-fabrication grounding rule and ships a runnable check (scripts/verify_citations.py; see faithfulness-and-citation.md) — use them, but they do not replace your own judgement.
• Never present unverified AI output as your own knowledge or as fact; do not submit AI-written text you have not checked and understood.
• Disclose your use of AI honestly, following your venue's and institution's policy — most journals and universities now require an AI-use statement (see declarations-guide.md).
• Follow your institution's academic-integrity rules.

Writing that reads like a scientist — and why that isn't a loophole

The writing skills actively remove AI writing tells: inflated significance ("plays a pivotal role", "paves the way"), vague attribution ("studies have shown"), stock vocabulary ("delve", "intricate landscape", "showcase"), "serves as" instead of "is", hedge stacking, and generic upbeat endings. These are bad scientific writing in their own right — vague, unattributed, and overstated — so cutting them makes a manuscript clearer, more specific, and more honest. See human-writing.md.

This is a quality edit, not a disguise. It is not a way to hide AI involvement, defeat an AI detector, or evade a journal's or university's AI policy. The AI-use disclosure requirement stands regardless of how the prose reads — the suite still writes that disclosure into every manuscript it drafts. The guide also refuses to trade accuracy for style: it keeps calibrated hedging (removing a "may" that the data require is a fabrication-grade error, not a polish), keeps passive voice where Methods need it, defers to the journal's house style, and never lets an edit change a value, a claim's scope, or a citation.

University of Melbourne policy (for UoM staff and students)

This project is initiated by the Food Science Research Team at the University of Melbourne; UoM users must comply with University policy on acknowledging and using generative AI. Always check the current policy and your course/coordinator's specific requirements:

• Acknowledging use of AI tools and technologies
• Academic integrity
• Using AI as a graduate researcher (graduate researchers and digital assistance tools)
• Guidelines for allowing student GenAI use in assessment
• Writing with GenAI
• Studying with GenAI
• Organising with GenAI
• Using GenAI effectively

Further reading

• Sarkar, R. (2026). Why AI can't be trusted to write scientific reviews. Nature. https://doi.org/10.1038/d41586-026-01616-3
• Using AI responsibly in scientific publishing. (2026). Nature Methods (editorial). https://doi.org/10.1038/s41592-026-03020-1

Citation

If you use these skills in your research, please cite the software:

Zhang, P., Liang, Z., Huang, P., Shen, C., & Yao, X. (2026). Academic Skills for Food & Nutrition Science [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.21372994

Zenodo mints two kinds of DOI — pick the one that matches what you mean:

State the version you actually ran in your methods or AI-use statement, whichever DOI you cite.

Machine-readable metadata is in CITATION.cff — GitHub's "Cite this repository" button generates APA and BibTeX from it. When you report AI-assisted work, also state the tool and model version you used (see Using AI responsibly and declarations-guide.md).

Contributing

We welcome contributions from food & nutrition research groups worldwide. If your team would like to contribute or collaborate, please contact the development team at [email protected].

Branching model — please read: main is release-only; never push to it directly. Do your work on the development branch and open a pull request to merge development → main, so changes are tracked and reviewed. Always keep README.md and CHANGELOG.md up to date in the same PR.

Full, machine-actionable instructions for collaborators and their AI coding agents are in AGENTS.md (see also CONTRIBUTING.md).

Key documents: README · CHANGELOG · LICENSE.

License & community

MIT — see LICENSE. Free for any use, including commercial. This open project was initiated by the Food Science Group, University of Melbourne (PangeZAU / PangenomeAI). Contributions from food & nutrition research groups worldwide are warmly welcomed — open an issue or pull request.

Contributors and institutions

• PangeZAU — Principal Investigator, Food Science Research Team, University of Melbourne.
• Zijian Liang — Research Coordinator, Food Science Research Team, University of Melbourne.
• Calebsch — Developer, Food Science Research Team, University of Melbourne.
• Pimiao Huang — Contributor, Food Science Research Team, University of Melbourne.
• Xiukun Yao — Contributor, Food Science Research Team, University of Melbourne.

Acknowledgements

This is original, independently written work released under MIT. It was informed by — but contains no code or text from — earlier community projects exploring academic-research and scientific-figure skills for Claude Code, including the nature-skills collection (Apache-2.0), deer-flow (MIT), Light-skills (MIT), academic-figure-skills and academic-figure-generator (MIT), Awesome-Journal-Skills (MIT), humanizer (MIT, Siqi Chen — whose AI-writing-pattern taxonomy, itself informed by Wikipedia's "Signs of AI writing", shaped our human-writing.md guide, rewritten for scientific manuscripts), and academic-research-skills (CC-BY-NC-4.0). Only non-copyrightable workflow concepts (e.g. multi-source search, layered retrieval, staged screening, parallel extraction, PRISMA structure, subagent teams, evidence/citation verification gates, AI-writing tells, and publication figure styles) were drawn on; all wording here is our own, so this project is free of their license obligations and is offered under MIT.