nature-skills
The academic-writing giant the scanners missed for five months finally gets a smoke test, and its figure QA tooling is the real thing.
This is a genuine engineering effort, not a prompt dump: six repo-level validators, eight CI workflows, a 100-test suite, and publication-figure QA scripts that we verified against both planted violations and real vendored PDFs. Install is byte-identical through the universal skills CLI for Hermes, and the demand is real at roughly 211K registry installs. Docked for a README that sells GPT recharge services before it explains installation, a frontmatter name rename that trips directory-name installs, and a heavy MCP dependency chain.
$npx skills add Yuan1z0825/nature-skills --list
$npx skills add Yuan1z0825/nature-skills --global --agent codex --skill '*' --yes --copy
$git clone https://github.com/Yuan1z0825/nature-skills.git ~/ai-skills/nature-skills
install if
- Graduate students and researchers drafting, polishing, or revising manuscripts who want the deepest publication-figure QA loop we have tested in a skill pack, with journal-family submission-requirement checks (Nature and NMI families in the test suite).
- Chinese-speaking academics: the docs, trigger words, and community channels are first-class Chinese with a CI-validated English mirror, and the paper2ppt plus response-to-reviewers workflows target CN lab conventions directly.
- Agent users outside Claude Code and Codex: the OpenClaw, OpenCode, and Hermes paths are documented, and the universal skills CLI install was verified byte-identical for hermes-agent, which is still unusual for packs this size.
skip if
- Anyone outside academic publishing: all 19 triggerable skills target the research paper lifecycle; there is nothing general-purpose here.
- Teams sensitive to commercial noise in dependencies: the README leads with a GPT/Claude recharge reseller and a paid community, and the DeepMind influence claim is unverifiable; the skills themselves are clean, but the project's presentation is sales-forward.
- English-only users wanting pinned releases: documentation is Chinese-first, five skills are labeled Draft, and there are no git tags or changelog to pin against, so you accept commit-level drift.
What It Does
nature-skills is a 20-directory research skill pack (19 triggerable skills plus the nature-shared support pack) covering the full academic publishing lifecycle: literature search with an embedded MCP server, paper reading and paper cards, citation retrieval and reference verification, manuscript writing and polishing, statistics review, reviewer simulation, response-to-reviewers, data availability statements, publication figures, paper-to-PPT and image-to-PPT conversion, experiment logging, a daily literature pipeline, and Chinese invention-patent drafting. The target user is a researcher (heavily Chinese-speaking, but with a validated English README mirror) driving Claude Code, Codex, OpenClaw, OpenCode, Hermes, or the Chatbox desktop app. Skills are prompt-contract SKILL.md routers backed by executable Python QA scripts, per-skill manifest.yaml routing axes, and progressive-disclosure references, static assets, and agents metadata. At review time the repo held 43,918 stars and 2,317 forks, was created 2026-04-24, and was pushed the morning of the review.
The Good
The publication-figure QA tooling is real software, and it passes both positive and negative tests. validate_figure.py runs 21 static checks (font family, colormaps, editable text, vector export presence, raster DPI, final width, demo-data boundaries, log guards, panel-alignment gate) on Python or R plotting source with zero required dependencies; matplotlib and PyMuPDF are imported lazily inside the functions that need them. We ran it on a vendored third-party example (14 pass, 4 warn, 3 fail) and on a planted violating script with a jet colormap, 5pt fonts, 72 dpi PNG export, and a 3.0 by 1.8 inch figure: it flagged all six planted violations by name (COLOR-MAP: Rainbow/jet/hsv color mapping is not publication-safe, RASTER-DPI: Raster resolution below 300 dpi, FONT-FAMILY, EDITABLE-TEXT, EXPORT-VECTOR, PANEL-ALIGNMENT-GATE) and exited 1 with verdict FIX BEFORE DELIVERY. audit_pdf_text.py extracts font sizes from a real vendored PDF using only stdlib zlib (minimum found 24pt against a 5pt floor, verdict PASS), and audit_figure_collisions.py shipped a stdlib-only --self-test that passes before PyMuPDF is ever imported.
Repository metadata is treated as software under test. Six validators ship in scripts/ and all pass: skill metadata (frontmatter keys, agents/openai.yaml presence for every triggerable skill, name-vs-manifest reconciliation, README badge count matching triggerable skills), skill index, per-skill READMEs, the CN/EN README mirror (20 mirrored docs), repository structure, and all eight GitHub workflow files. Eight CI workflows run content-contract tests (17 test files covering abstract guidance, figure collision audit, submission requirements per journal family) and skill-tooling tests. The local suite is 100 passed, 9 skipped, 8 subtests in 0.10 seconds on a 42MB venv containing only PyYAML and pytest, and the two shell test suites for the Codex sync and auto-update scripts both pass (state isolation, retry, locking; prune safety with transactional rollback).
License hygiene is more careful than most of the tier. The root is Apache-2.0, and skills/nature-figure/assets/figures4papers/THIRD_PARTY_NOTICES.md names the upstream project (ChenLiu-1996/figures4papers), retains upstream copyright, and explicitly states the vendored files must not be assumed to fall under the repo license because upstream exposed none as of the checked date. The figure validator docstring also declares its lineage: it incorporates and refactors ideas from an Apache-2.0 academic-figure-skill workflow.
The universal install path is verified clean, and the demand is measurable. npx skills add found all 20 skills from a local clone, and a full --skill '*' install for hermes-agent landed 20 directories, 20 SKILL.md files, and 747 files total, byte-identical to the source on spot diff (nature-figure and nature-shared both matched; the only diff entry was a local __pycache__ the installer correctly omits). A scoped two-skill install (nature-reader plus nature-shared) landed identically with a skills-lock.json carrying SHA-256 computed hashes. The skills.sh registry carries 21 canonical entries for this owner totaling roughly 211,000 installs (nature-figure 15,204; nature-polishing 14,019; nature-paper2ppt 12,900).
Honest state labeling and fail-closed behavior throughout. The skill index stamps each skill Stable, Beta, or Draft (4 Stable, 10 Beta, 5 Draft at review time). Every dependency-guarded script fails closed with an actionable message: the metadata validator exits 1 with Missing dependency: PyYAML. Install with python -m pip install pyyaml., the backend state script documents and implements exit-1-when-unset semantics, and the MCP server refuses to start without its modules rather than half-booting.
The Bad
The README sells GPT and Claude account recharge services before it explains installation. Sections 1.2 through 1.4 of the primary README promote a self-operated recharge card site (apiciyuan.top) for GPT plus and Claude Pro top-ups, a paid Knowledge Planet community, WeChat customer-service QR codes, and Douyin tutorial channels, all ahead of the quick start. For a 43.9K-star Apache-2.0 repo, the first screen of documentation is an ad for a gray-market-adjacent reselling service. Section 3.1 also claims Google DeepMind borrowed the citation system and design philosophy for their own Science Skills; the claim links to nothing and cannot be verified.
One skill's frontmatter name does not match its directory, and the fix is documentation rather than repair. Directory nature-proposal-writer carries frontmatter name researchwrite, so npx skills add --skill nature-proposal-writer fails while --skill researchwrite succeeds; the registry carries the rename (9,005 installs under researchwrite). The repo validator codifies it as a COMPATIBILITY_SKILL_NAMES exception instead of migrating the name, and the README asks users to remember the mapping. Anyone guessing directory names (the natural convention across the other 19 skills) hits an error.
The registry carries a ghost entry. skills.sh lists openclaw-medical-skills under this owner with 255 installs, but no such directory exists in the current repo; the 20 on-disk skills are all nature-prefixed. It is retirement drift, the same defect class we flagged in vendor packs ten times this size.
The heaviest dependency chains gate the most useful integrations. The nature-academic-search MCP server imports pybliometrics at module scope through its sources package, so the server cannot even reach the tool-list handshake without installing the Scopus SDK plus its pandas stack (our review budget caps installs at 100MB, so the handshake went unverified). The repo's own CI additionally installs libreoffice-impress and playwright chromium for the image2ppt and patent-disclosure paths. None of this is hidden; all of it is opt-in per skill, but the pack's advertised flagship search surface is its most expensive install.
No releases, no tags, no changelog, and version drift is structural. The only version string in the repo is 0.1.0 inside .codex-plugin/plugin.json; there are no git tags and no CHANGELOG, so consumers diff commits to track changes. Minor hygiene nits stack up: the root tests/ directory contains a duplicate basename (test_generate_star_history.py, same file as scripts/tests/) that breaks pytest collection when both directories are passed together, and documentation is Chinese-first with the English mirror validated but secondary.
Smoke Test Results
Host-based hands-on testing on macOS (aarch64), 2026-09-22. Fresh runs used the system Python 3.9 with no scientific packages installed; dependency runs used a Python 3.11 venv. The sbx daemon was not authenticated this session, so runs were recorded on the host with logs captured to the standard sandbox paths.
Run A. Fresh sandbox, no Python deps preinstalled
$ git clone --depth 1 https://github.com/Yuan1z0825/nature-skills.git /tmp/nature-skills-fresh
Cloning into '/tmp/nature-skills-fresh'... (78M)
β
clone completes; 78MB working tree
$ find skills -name SKILL.md | wc -l && find skills -maxdepth 1 -mindepth 1 -type d | grep -cv nature-shared$
20 / 19
β
20 SKILL.md files: 19 triggerable dirs plus nature-shared support pack
$ python3 scripts/validate-skill-metadata.py
Missing dependency: PyYAML. Install with `python -m pip install pyyaml`.
exit=1
β
fails closed with the documented actionable message (system python has no pyyaml)
$ python3 skills/nature-figure/scripts/nature_figure_backend.py get; echo "exit=$?"
exit=1
β
documented semantics: exit 1 when no default is saved, no output
$ python3 skills/nature-figure/scripts/audit_figure_collisions.py --self-test
audit_figure_collisions.py self-test: PASS (exit=0)
β
stdlib-only geometry core passes without PyMuPDF
$ npx -y skills add /tmp/nature-skills-fresh --list | grep -E "Found|validated"
β Local path validated
β Found 20 skills
β
universal installer discovers all 20 skills from a local path, no Python deps needed
$ python3 ns_frontmatter_scan.py /tmp/nature-skills-fresh
name drift count: 1
DRIFT: dir=nature-proposal-writer frontmatter=researchwrite
β
scan completes; exactly one name drift, matching the README's disclosure
$ ls LICENSE
LICENSE (Apache-2.0, API-confirmed spdx apache-2.0)
β
license file present
Pass rate: 8 of 8. Fresh-clone behavior matches every documented contract, including the two deliberate fail-closed paths.
Run B. Sandbox with deps preinstalled
$ /Users/openclaw/.local/bin/python3.11 -m venv /tmp/ns-venv && pip install pyyaml pytest
42M venv (Python 3.11.14)
β
tooling deps install in a 42MB venv
$ for v in validate-skill-metadata validate-skill-index validate-readmes validate-readme-mirror validate-repository validate-workflows; do python scripts/$v.py; done
Skill metadata validation passed: 20 skill directories, 19 triggerable skills.
Skill index validation passed: 19 triggerable skills, 20 listed entries.
README validation passed. / README mirror validation passed: 20 mirrored skill docs checked.
Repository validation passed: 20 skill directories checked. / Workflow validation passed: 8 workflow files checked.
β
all six repo validators exit 0
$ python -m pytest scripts/tests -q
100 passed, 9 skipped, 8 subtests passed in 0.10s
β
full tooling test suite green (root tests/ has a duplicate basename that breaks combined-dir collection)
$ bash scripts/tests/test-autoupdate-state-isolation.sh
Auto-update state isolation, retry, and locking passed.
β
Claude Code auto-update script tests pass
$ bash scripts/tests/test-update-codex-skills-safety.sh
Skill prune safety and transactional rollback passed.
β
Codex sync script tests pass
$ npx -y skills add /tmp/nature-skills-fresh --agent hermes-agent --skill '*' --yes --copy
20 copied lines; landed skill dirs: 20; landed SKILL.md: 20; landed files total: 747
β
full 20-skill install lands 747 files for hermes-agent
$ diff -r -x __pycache__ skills/nature-figure $PROJ/.hermes/skills/nature-figure && diff -r skills/nature-shared $PROJ/.hermes/skills/nature-shared
nature-figure IDENTICAL (excl local __pycache__) / nature-shared IDENTICAL
β
installed copies byte-identical to source
Pass rate: 7 of 7. With only PyYAML and pytest, every repo-owned gate turns green and the complete pack installs byte-identical through the universal installer.
Run C. Functional verification (does it do what it claims?)
$ BK=nature_figure_backend.py; HOME=/tmp/ns-func-home python $BK set python && python $BK get && python $BK path && python $BK clear && python $BK get; echo "exit=$?"
python / /tmp/ns-func-home/.config/nature-skills/nature-figure.json / get-after-clear exit=1
β
state machine roundtrip correct: set saves, get reads, path resolves, clear resets to exit-1-unset
$ python validate_figure.py --backend python assets/figures4papers/figure_Dispersion/plot_illustration.py
summary: 14 pass, 4 warn, 3 fail / verdict: FIX BEFORE DELIVERY (exit=1)
β
positive control: the tool runs on a vendored third-party script and honestly fails it (no panel-alignment gate call, 4 add_subplot calls)
$ python validate_figure.py --backend python /tmp/ns-func-out/planted_bad_figure.py
[FAIL] FONT-FAMILY / COLOR-MAP: Rainbow/jet/hsv / EDITABLE-TEXT / EXPORT-VECTOR / RASTER-DPI: below 300 dpi / PANEL-ALIGNMENT-GATE
summary: 12 pass, 3 warn, 6 fail / verdict: FIX BEFORE DELIVERY
β
negative control: all six planted violations flagged by name (jet colormap, 5pt fonts, 72 dpi, 3.0x1.8in figure, PNG-only export, no alignment gate)
$ python audit_pdf_text.py assets/figures4papers/figure_Cflows/figures/fig2_comparison_GeneRegulatory.pdf
minimum required: 5 pt / minimum found: 24 pt / text runs: 42 / below minimum: 0 / verdict: PASS
β
real PDF audit: stdlib zlib extraction, correct floor enforcement
$ python audit_figure_collisions.py assets/figures4papers/figure_Cflows/figures/fig2_comparison_GeneRegulatory.pdf
summary: 3 fail, 6 warn, 0 contained fill overlays / verdict: FIX BEFORE DELIVERY
β
collision auditor finds real text-over-fill-edge overlaps in the vendored figure (PyMuPDF installed for this run)
$ cd skills/nature-academic-search/mcp-server && python academic_search_server.py
ModuleNotFoundError: No module named 'mcp' (exit=1) / tool registrations in source: 16
β
MCP server fails closed without deps; 16 @mcp.tool registrations counted in source (handshake not run: pybliometrics chain exceeds the 100MB install budget)
$ curl -s "https://skills.sh/api/search?q=nature-" | python /tmp/parse_skillssh.py
21 canonical owner entries, ~211,332 total installs / incl. ghost openclaw-medical-skills (255 installs, no such dir in repo) and rename entry researchwrite (9,005)
β
registry probe confirms canonical entries, demand, and both drift findings
Functional pass rate: 7 of 7. Every executable claim we could reach within budget was verified with a positive and a negative case.
What the runs tell you
The executable layer does exactly what its documentation promises, including failing the repo's own vendored showcase examples for missing alignment gates. The two unverifiable surfaces are the MCP tool calls (blocked by the pybliometrics install weight, not by any defect we observed) and the in-agent prompt-contract workflows, which need a live session by design.
Setup Walkthrough
- Clone or point the installer at the repo:
npx skills add Yuan1z0825/nature-skills --listshows all 20 installable names. Note thatnature-proposal-writerinstalls asresearchwrite. - Install scoped or full for your agent:
npx skills add Yuan1z0825/nature-skills --global --agent codex --skill nature-figure --yes --copy, or--skill '*'for everything. Add--skill nature-sharedalongside single doc-skill installs; the README documents which skills depend on it. - Hermes users have two documented paths: copy skill directories into
~/.hermes/skills/, or declare the repo'sskills/directory as an external skill source in~/.hermes/config.yamlperdocs/open-source-agent-frameworks.md. Our hermes-agent CLI install landed at project level (./.hermes/skills/) byte-identical. - Claude Code users keep a stable clone and either write the README's subagent or slash-command wrappers pointing at the real SKILL.md paths, or run
scripts/autoupdate-skills.sh --forceonce and add the documented SessionStart hook (async, 1-hour network throttle, skip-on-failure by design). - Optional runtimes are per-skill and opt-in: PyYAML for repo validators, matplotlib or ggplot2 for figure rendering, PyMuPDF for collision audits, the pybliometrics chain for the academic-search MCP server, playwright chromium for CNIPA patent searches.
Post-install gotcha: skills are routed by their bilingual trigger descriptions, so English-only prompts route fine but the Chinese trigger words are the primary ones for several skills.
Alternatives
- WenyuChiou/ai-research-skills - a registry-style catalog routing to 17 skills across five upstream repos with a 3-gate thesis filter; prefer it when you want curated multi-repo coverage over one org's monorepo with its own QA gates.
- yjz211/vivid-figures-skill - the CN scientific-plotting alternative with 108 SHA-pinned recipe wrappers around a plotting suite; figures only, non-commercial license, where nature-figure is Apache-2.0 with static analysis instead of pinned recipes.
- blader/humanizer - a doc-only de-AI writing contract; complements nature-polishing as a final style pass rather than replacing the academic rewrite workflow.
Reviews stay honest because nobody pays us to publish them. If this one saved you time, throw a coin.
Tip the reviewer- reviewed by
- GearScope
- tested
- 2026-09-22 · macOS (Apple Silicon)
- last verified
- 2026-09-22
- depth
- HANDS-ON
- sponsorship
- none, ever
Don't install your next skill blind.
Every week: the shortlist of skills worth installing β and the ones to skip β from 100+ hands-on tests. No spam, no affiliate links.