Qinyan Academic Skills: 187 Agent Skills for Academic Research, Installed by Shell Script
A curated, multilingual library of 182 installable AI agent skills for end-to-end academic research—spanning literature discovery, scientific writing, grant development, bioinformatics, drug discovery, clinical research, machine learning, and data analysis.
At a glance
- What is it?
- A curated library of installable AI agent skills covering literature search, Nature-style manuscript work, bioinformatics and drug discovery, distributed as plain text and installed with a Bash script. The useful question is whether prompt-level workflow guidance is the right abstraction for your research pipeline.
- Who is it for?
- Adopt it if you already work inside Claude Code, Cursor, Codex, Gemini CLI, OpenClaw or OpenCode and want curated, reviewable prompt workflows for literature search, manuscript drafting or bioinformatics without writing them yourself. Skip it if you need deterministic, testable analysis code, because a skill is guidance the agent may or may not follow, not a function you can assert on.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 57 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the Repository Actually Ships
The README describes 187 skills across 18 domains, distributed as plain text files under skills/, with a five-skill first-party suite for Nature-style scientific writing. The catalog table lists categories by ID: literature discovery and reference management (10 skills), scientific writing and communication (6), academic presentation and visualization (9), research methods and scientific reasoning (10), bioinformatics and genomics (21), and cheminformatics and drug discovery (12), with the remainder not shown in the excerpt. A separate Qinyan Academic skills group holds 10 skills covering paper search, analysis, citations, and the Nature-focused writing, review, figures and statistics workflows. The description supplied with the repository spans literature discovery, grant development, bioinformatics, drug discovery, clinical research, machine learning and data analysis, which matches the category names visible in the README. One inconsistency is worth noting: the repository description says 182 skills, the README badge and prose say 187. Treat the catalog itself as the source of truth and count what you install rather than trusting either number. The licence is MIT, and the README states each skill is plain text, reviewable and portable, which is the property that makes the selective installation story credible.
The Installer Is the Product Boundary
Everything in this repository funnels through a single install.sh fetched over curl. The README gives these forms: piping the script with no arguments installs the full library into the global Claude Code skills directory; --category 01 installs one category; --skill scanpy or --skill qinyan-nature-writing installs a single named skill; --project redirects installation into the current project instead of the global directory; --tool codex targets a different agent. Short forms -c, -s and -t are documented as equivalent, and bash install.sh --help prints the full reference. Maintenance commands are part of the same script: --list and --list-skills browse the catalog, --search "protein" filters by name and description, --status shows what is installed, --check-update and --update handle upgrades, and --update --skill scanpy narrows an update to one skill. The installer requires Bash, Git and curl, and the README tells Windows users to run it under WSL or Git Bash. It also suggests reviewing install.sh before piping it to Bash, which is the right instinct for a script that writes into your home directory.
Agent Portability and Where Files Land
The supported-agent table is the most concrete part of the documentation. Six agents are listed with a --tool value and two directories each. Claude Code (claude, the default) uses ~/.claude/skills/ globally and .claude/skills/ per project. Cursor uses ~/.cursor/skills/ and .cursor/skills/. Codex uses ~/.codex/skills/ and .codex/skills/. Gemini CLI uses ~/.gemini/skills/ and .gemini/skills/. OpenClaw uses ~/.openclaw/skills/ and .openclaw/skills/. OpenCode is the outlier: its global path is ~/.config/opencode/skills/ while its project path is .opencode/skills/, so a global install for OpenCode does not mirror the pattern the other five follow. If you keep one skills directory under version control and symlink it into each agent's expected location, that asymmetry will trip you up the first time. The portability claim is really a claim about file layout: the same skill text is copied into whichever directory the chosen agent scans. Nothing in the material suggests the skill format is translated or adapted per agent, so behavioural differences between agents running the same skill are not addressed by the installer.
The Nature-Style Suite and Its Five Stages
The five first-party skills are staged across a manuscript lifecycle, and the README assigns each a primary output. qinyan-nature-writing covers argument and drafting, producing an evidence-led narrative, section drafts, title, abstract and submission package. qinyan-nature-polishing handles structural and language revision, with what the README calls meaning-preservation and anti-overclaim checks. qinyan-nature-review performs pre-submission assessment and returns prioritized, traceable findings with stable issue identifiers, a detail that matters if you want to re-run the review after edits and confirm that a specific finding was actually resolved. qinyan-nature-figures targets multi-panel figures with export and preflight validation. qinyan-nature-statistics produces estimand-aware analysis plans, reporting audits and figure-ready results. The estimand language is a specific choice: it implies the statistics skill asks what quantity is being estimated before choosing a method, which is the correct order and is frequently skipped. What the README does not provide is the actual text of any skill, the quality gates referenced in NATURE-SKILLS.md, or example prompts beyond the pointer to that file. The disclaimer is explicit and worth repeating: Nature-style describes editorial and scientific communication goals, and the project states it is not affiliated with, endorsed by, or guaranteed acceptance by Nature Portfolio or Springer Nature.
A Skill Is Guidance, Not a Pipeline
The central limitation is structural. A skill in this repository is a text artifact placed in a directory the agent reads. It shapes what the model does; it does not execute a fixed sequence, validate inputs, or fail loudly when a step is skipped. For literature search that is usually acceptable, because the cost of a mediocre search is a wasted afternoon. For the bioinformatics and cheminformatics categories, which the catalog lists at 21 and 12 skills respectively, the same looseness is a real risk. If a skill wraps scanpy for single-cell analysis, the agent still writes the code, and the reproducibility of the result depends on the model's choices at generation time rather than on a pinned version of anything. Nothing in the supplied material documents pinned dependency versions, test suites, or output validation for those skills. The statistics skill's reporting audits have the same character: an audit performed by a language model is a review, not a check. If your work needs a result you can re-derive byte for byte, this repository is the wrong layer and a scripted Snakemake or Nextflow pipeline is the right one.
How It Compares to a Fixed Prompt Library
The obvious alternative is maintaining your own prompts and snippets in a dotfiles repository or a shared team folder. That approach gives you total control over content and costs nothing to set up, but you own the catalog problem: discovering that a better literature-review prompt exists, keeping it current as models change, and distributing it consistently across six different agents. Qinyan's contribution is not a novel technique, it is the packaging. The installer is the differentiator, because it resolves one name against a curated catalog and writes to the correct directory for the agent you named. The trade-off is trust and coupling. Piping a remote script to Bash means the project controls what lands in ~/.claude/skills/, and the --update path means that content can change under you without a version pin. The README offers no release tags in the material provided, so you cannot pin to a specific revision through the documented commands. Cloning the repository with git clone and running bash install.sh from the checkout removes the remote-script risk and lets you diff changes yourself, which is the safer route if you intend to depend on these skills for published work.
Maintenance Cost and Licence Position
Upgrade cost is low by design. --check-update reports available changes and --update applies them, optionally scoped with --skill. Because skills are plain text, a git diff of your installed directory shows exactly what changed between versions, which is more than most prompt collections offer. The cost that remains is review time: 187 skills is far too many to read, so the practical approach is to install narrowly with --skill or --category and inspect only what you use. The MIT licence permits commercial use, modification and redistribution, and the repository ships a LICENSE file at its root. That covers the project's own text. It does not cover anything a skill instructs an agent to fetch or call, such as literature databases or third-party Python packages, and those carry their own terms. This is a description of the licence, not legal advice; if you redistribute a modified skill set inside a commercial product, read the LICENSE file and your own obligations rather than relying on a summary. The last push date in the supplied metadata is 2026-07-20, and no releases are listed, so versioning appears to be commit-based.
Editorial conclusion
Adopt it if you already work inside Claude Code, Cursor, Codex, Gemini CLI, OpenClaw or OpenCode and want curated, reviewable prompt workflows for literature search, manuscript drafting or bioinformatics without writing them yourself. Skip it if you need deterministic, testable analysis code, because a skill is guidance the agent may or may not follow, not a function you can assert on. Before installing anything, run bash install.sh --list-skills and read the SKILL.md of the exact skill you intend to use, then run the installer with --project rather than piping it globally, so the files land in .claude/skills/ where you can inspect and remove them.
Community notes