How to use the Nature Reference Verifier agent skill

What the skill makes the agent do

The skill verifies existing bibliography entries; it does not write new references. For each entry, the agent first checks the DOI against Crossref (api.crossref.org). If the DOI does not resolve, it is treated as a likely wrong number. If it resolves, the agent compares the returned record with your entry.

The agent then compares authors, title, year, volume, issue and pages against other sources such as IEEE Xplore, web search, and CNKI/Wanfang for Chinese literature. Mismatches are graded as critical (must fix), warning (check) or info (for reference only).

• Critical: wrong or reordered authors, missing authors, page differences of 5 or more, letter-number confusion in article numbers, DOI pointing to a different paper
• Warning: volume year differing from DOI year, Early Access metadata that changed after print, missing issue, small page shifts, publisher name variants
• Info: title capitalisation, abbreviated vs full journal names, punctuation differences

Install the skill from the ZIP

Download the ZIP of the skill from this page and unzip it. The skill folder must end up in a skills directory where your agent looks for skills. For Claude Code, that is ~/.claude/skills/ for all your projects, or .claude/skills/ inside a single project.

If you use the Claude apps instead, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal (all projects): unzip into ~/.claude/skills/
• Single project: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with a natural request

The skill activates on phrases such as "verify references", "check references", "ref check", "校验文献" or "核对参考文献". You can also describe the job in plain language. The project's own examples include asking the agent to check 50 references one by one and mark the ones that must be fixed, or to find the real titles and page ranges for DOIs you suspect are wrong.

Give the agent the input in one of three forms: a full reference list from a paper or proposal (numbered [N] entries, with DOIs where they exist), a single citation pasted as text or a Zotero item key, or a BibTeX file.

What the report looks like

Every reference gets one of four statuses: Verified (sources agree), Check suggested (warning-level differences need your judgement), Needs fix (critical differences must be corrected), or Unverifiable (no source could find it, for example an internal report or an old thesis).

The report can be a Markdown summary with tables of must-fix items and items to check, a BibTeX patch with corrected fields, or Zotero update instructions. Each finding shows the current value, the value from the source, and where it came from.

Updating Zotero safely

The skill's notes say the local Zotero 7 HTTP API (localhost:23119) is read-only, so writes there fail. For bulk corrections the skill prefers pyzotero in Web mode, which needs a Zotero Web API user ID and API key. Without a key, the local connector endpoint can add new items but cannot change existing ones. Editing zotero.sqlite directly is described as a last resort, only with Zotero closed.

Before changing anything, confirm the entry is actually in your Zotero library. If a title search returns nothing, the bibliography likely came from another source such as a CSV, and Zotero needs no change.

Safety, network and what you need

The skill does not run scripts. It does use the network: it queries Crossref and performs web searches, and it can use optional sources such as IEEE Xplore, CNKI/Wanfang through kimi-webbridge, kimi-datasource scholar search, pyzotero with your Zotero Web API credentials, or a local Zotero 7 for read-only lookups. Without the optional tools it falls back to basic web search.

Treat your Zotero API key as a secret, and only give the agent sources you are allowed to use. The skill itself notes that some differences need manual judgement, so review every Needs fix and Check suggested item before you submit.

• Requires internet access
• No scripts are run by this skill
• Optional credentials (Zotero Web API user ID and key) are only needed for writing corrections

Tips for better results

Send long lists in batches. The skill's guidance is to split more than 20 entries into groups of 10 to 15 and have parallel sub-agents check each group, then combine the results. Check DOIs first, because a DOI that returns 404 or points to another paper is the cheapest error to find.

Expect gaps for Chinese theses without DOIs, unpublished reports, and older IEEE records; these often come back Unverifiable rather than wrong.

• Start with the DOI check: it catches many errors in one request
• Keep the original file so you can compare the patch with what you had
• Use the skill on the reference list only; it is not a writing tool