Skip to content
Get packs

Guide · Agent skill

How to use the Citation Audit agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The citation-audit skill is a set of instructions and scripts that lets an AI coding agent check the bibliography of a LaTeX paper before you submit it. It checks that each cited work exists, that its authors, year and venue are correct, and that the cited paper actually supports the sentence where it is used. This guide explains what the skill does, how to install it, how to trigger it, what it needs, and what to watch for when it runs.

Citation Auditwanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Download the ZIP. Download the citation-audit ZIP from the skill page on this site.
  2. 2Unzip into a skills folder. Unzip it into ~/.claude/skills/ for all projects, or into .claude/skills/ inside one project.
  3. 3Check your reviewer and search tools. Make sure a reviewer backend is configured (default Codex MCP with gpt-6-astra) and that web search and fetch are available to the agent.
  4. 4Ask for the audit once the bib file is final. Say something like 'check citations' or 'citation audit' after paper-writing and claim checks are done and before final compile.
  5. 5Review the report. Open CITATION_AUDIT.md and CITATION_AUDIT.json, check each verdict against the source yourself, and decide what to change.

What the citation-audit skill makes the agent do

The skill checks every bibliographic entry that your paper cites against three layers. The first is existence: the cited paper must actually exist at the claimed arXiv ID, DOI or venue. The second is metadata correctness: author names, year, venue and title should match canonical sources such as DBLP, arXiv, ACL Anthology, Nature or OpenReview. The third is context appropriateness: the cited paper should support the claim it is used to support in your manuscript.

The skill targets the problems that are hard to notice by eye, such as a real paper cited for a claim it does not make, anonymous-author placeholders that slipped through, arXiv v1 and v3 titles silently merged, a preprint cited when the official conference version exists, and DOIs that do not resolve. The output is a per-entry report named CITATION_AUDIT.md and a machine-readable JSON ledger named CITATION_AUDIT.json.

Install the citation-audit skill

Option one: download the ZIP for citation-audit from the skill page, then unzip it into your personal skills folder at ~/.claude/skills/. To make it available only to one project, unzip it into the .claude/skills/ folder inside that project instead.

Option two: if you use the Claude apps, upload the ZIP there under Settings → Capabilities → Skills.

  • Personal install location: ~/.claude/skills/
  • Project install location: .claude/skills/ inside your project
  • Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Trigger the skill with a natural request

You do not need a special command. The skill is picked up when you ask for a citation check in words like the ones below. You can also say you want the audit run before submission, and the agent should use this skill for that task.

The skill's argument hint is a paper directory or bib file, optionally followed by --uncited or a soft-only flag. Those options are described in the next section.

  • check citations
  • citation audit
  • verify references
  • 引用核对
  • 审查引用
  • Check the citations in my paper before I submit it.
  • Run a citation audit on references.bib and the sections folder.
[paper-directory-or-bib-file] [--uncited] [— soft-only]

What the skill needs before it runs

The skill expects a LaTeX paper with a references.bib file (or similar bib file) and \cite commands in your text files, typically in a sec/ or sections/ folder. It also needs a configured reviewer backend. The default route is Codex MCP with the model gpt-6-astra. The skill also needs web search and fetch tools so it can look up DBLP, arXiv, ACL Anthology and similar sources.

The skill is written to run after your bibliography is final. It is not meant for a draft that still contains placeholder text, because the context check depends on the surrounding claims being settled.

  • A LaTeX paper with references.bib and \cite commands
  • A configured reviewer backend (default: Codex MCP, model gpt-6-astra)
  • Web search and fetch tools for DBLP, arXiv, ACL Anthology and similar lookups

Run the audit and read the output

The agent first finds your bib file and the .tex files that contain \cite calls. It records each cite key with its file, line number and surrounding sentence, and saves those contexts to .aris/citation-audit/contexts.txt inside your paper directory. Then it sends each cited entry to a fresh reviewer context, so no entry is judged in the same thread as another.

Only cited entries are audited by default. Bib entries that are never cited are not sent to the reviewer. If you add --uncited, the skill also lists bib entries that are in the file but not cited anywhere in the paper.

When the audit finishes, the agent writes CITATION_AUDIT.md for people to read and CITATION_AUDIT.json for tools to read. Review every verdict yourself, because the verdicts come from an AI reviewer that can still make mistakes.

  • Default: audit cited entries only, with a fresh reviewer context per entry
  • --uncited: also list bib entries that are never cited
  • Output: CITATION_AUDIT.md and CITATION_AUDIT.json

Safety notes: scripts, network and bib edits

The skill runs scripts and uses the network. It looks up sources online and its allowed tools include shell commands, file read and write, editing, and web search and fetch. Read the SKILL.md file before you run it on a paper you care about, and make sure you have a backup or version control for your bib file.

By default the skill can change the bib file. Findings that call for a fix, replacement or removal are applied to references.bib. If you want the bib file frozen, use the soft-only flag. In soft-only mode the skill is forbidden to mutate any .bib file, and it instead proposes sentence rewrites in the citing .tex files, one per occurrence.

  • Default mode may FIX, REPLACE or REMOVE bib entries
  • Soft-only mode never mutates the bib file and proposes sentence rewrites instead
  • The reviewer is run with a read-only sandbox setting in the workflow
  • Keep a backup or version control before running

What people use it for

  • Catching a real paper that is cited for a claim it does not establish before a conference or journal submission.
  • Finding a hallucinated author list or wrong year in a thesis reference list before handing it in.
  • Checking whether an arXiv preprint should be replaced by its official venue record.
  • Running a repeatable bibliography check as one gate in a paper-writing pipeline, after the bib file is finalized.
  • Using soft-only mode to get sentence-level rewrite proposals while keeping references.bib untouched.

Frequently asked questions

Is the citation-audit skill free to use?

The skill's source is published under the MIT licence, as listed on the skill page. The reviewer backend and any model or search service you use may have their own terms, which this page does not cover.

What do I need to install the citation-audit skill?

You need to download the ZIP from the skill page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it in the Claude apps under Settings → Capabilities → Skills. The audit also needs a configured reviewer backend and web search and fetch tools.

Does the citation-audit skill work with any LaTeX paper?

It is written for a LaTeX paper with a references.bib file and \cite commands. The default setup uses Codex MCP with the model gpt-6-astra as the reviewer.

Why does the audit say the reviewer is unavailable?

The skill sends each reviewer slot through the configured reviewer route. If the selected reviewer cannot be reached, the slot is marked REVIEW_UNAVAILABLE rather than being filled with a guess. Check that your reviewer backend is set up and that the agent has web search and fetch access.

Can the citation-audit skill edit my references.bib file?

Yes, by default it may fix, replace or remove bib entries. If you pass the soft-only flag, it is not allowed to change the bib file and proposes sentence rewrites in the .tex files instead.

How is this different from just running a search for each reference?

The skill also checks context, meaning whether the cited paper supports the sentence where it is used, not only whether the paper exists. Its verdicts still come from an AI reviewer, so a human should make the final call.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use citation-audit if you are a researcher or student preparing a finished paper or thesis and want a second, repeatable check of your references and their context before submission. Pick something else if your draft still has placeholder text, if you need a fully automatic gate that runs on a timer, or if you cannot allow a skill to run scripts and use the network.

Download the skill