Skip to content
Get packs

Guide · Agent skill

How to use the Novelty Check agent skill

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

The Novelty Check skill is an instruction package for an AI coding agent, such as Claude Code, that checks whether a research idea has already been done in the published literature. It is for researchers and students who want a structured first pass on prior work before they invest time in a pilot experiment or a proposal. By the end of this guide you will know what the skill does, how to install it, how to trigger it with a plain request, and what it needs to run.

Novelty Checkwanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Download and unzip. Download the skill ZIP from this page and unzip it so that the novelty-check folder with SKILL.md is available.
  2. 2Place the skill folder. Copy the folder to ~/.claude/skills/ for all projects, or to .claude/skills/ inside one project.
  3. 3Start the claude-review bridge. Make sure the local claude-review MCP bridge is running, since the reviewer step depends on it.
  4. 4Describe your idea. Write a short description of the method, the problem it solves and what makes it different from obvious baselines.
  5. 5Ask for a novelty check. Use a trigger phrase such as novelty check or 有没有人做过 so the agent starts the search and review.
  6. 6Review the report and trace. Read the score, recommendation and closest papers, then check the saved trace in .aris/traces/novelty-check/ before deciding.

What are agent skills and what does this one do?

An agent skill is a folder with a SKILL.md file that tells an AI agent how to carry out a specific task. When the skill is installed and triggered, the agent follows its steps instead of improvising.

This skill tells the agent to take a research idea and break it into three to five core technical claims. For each claim it searches arXiv, Google Scholar and Semantic Scholar, plus recent conference papers, and reads the abstracts of the closest matches. It then sends the idea and the list of papers to a separate Claude reviewer through a local claude-review MCP bridge, which judges novelty.

  • Output: a structured Novelty Check Report with a score out of 10, a PROCEED, PROCEED WITH CAUTION or ABANDON recommendation, and a one-sentence statement of what is new.
  • Trace: the search and review are saved to a project folder at .aris/traces/novelty-check/ so you can audit how the verdict was reached.
  • Verdict rules: the skill tells the reviewer that proximity to other work is not a reason to reject, and that ABANDON requires naming a specific published paper that already contains the result.

Install the Novelty Check skill

Download the ZIP from this page and unzip it. Then choose one of two options. For a personal setup that works in all projects, place the skill folder in ~/.claude/skills/. For a single project, place it in the project's .claude/skills/ folder.

If you use the Claude apps rather than Claude Code, you can upload the skill instead. Go to Settings, then Capabilities, then Skills, and upload the package there.

If you use Codex and want Claude to act as the reviewer, the source says to install this package after skills/skills-codex/*, because it overrides that setup.

  • Personal install: ~/.claude/skills/
  • Project install: .claude/skills/ inside your project folder
  • Claude apps: Settings → Capabilities → Skills

Set up the reviewer requirements

The skill cannot finish its reviewer step on its own. It needs the claude-review MCP bridge, which is a local MCP server, and an agent with WebSearch and WebFetch tools. Without those, the search and review steps described in the skill will not run as written.

The CLAUDE_REVIEW_MODEL environment variable is optional. Set it only if you want to override the Claude model used by the reviewer. The oracle-pro reviewer is also optional and is used for additional review rounds.

  • Required: Claude Code or a compatible agent with WebSearch and WebFetch
  • Required: claude-review MCP bridge running locally
  • Optional: CLAUDE_REVIEW_MODEL to choose a specific Claude model
  • Optional: oracle-pro reviewer for extra review rounds

Trigger the skill with natural requests

You do not need a special command. The skill is meant to activate when you ask whether an idea is new, using phrases such as the ones listed in its description. Describe the method in a few sentences, including the problem it solves and what makes it different from obvious baselines. The more specific the description, the better the claim extraction.

The skill's description lists these trigger phrases: 查新, novelty check, 有没有人做过, check novelty, and requests to verify that a research idea is novel before implementing it.

  • "Do a novelty check on this idea: ..."
  • "Check novelty: I want to use retrieval to ..."
  • "有没有人做过 ..."
  • "Verify this research idea is novel before I start implementing it."

Read the report and the trace

The report starts with the proposed method and lists each core claim with its closest paper and the difference that remains. A table of closest prior work follows, then an overall assessment with the score, the recommendation, the key differentiator and the risk a reviewer would raise.

The skill's score anchor is 5 out of 10, which means the idea has clear neighbors but a defensible difference worth a pilot. Scores of 1 to 3 are reserved for results that a named published paper already contains. The trace folder records the searched claims, closest papers, reviewer route and raw reviewer response, so you can check the reasoning.

  • PROCEED: crowded ground with a clear, verifiable difference
  • PROCEED WITH CAUTION: the difference is thin or contested, and the report says what would make it carry
  • ABANDON: a specific published paper already contains the result, and that paper is named
review_independence: cross-family acceptance_status: accepted

Use the skill responsibly

The skill itself calls this an early gate, not a final verdict. It can miss papers or misjudge them, so a human literature review is still needed before publication. Treat the result as a reason to look closer, not as proof of novelty.

The skill relies on web search and an external reviewer bridge, so your idea description and review content leave your machine through those services. Do not use it if your project cannot send search queries or review content to external services. It is also designed for judging the academic novelty of methods, so it is not suited to non-research tasks.

  • Scripts: the source lists no scripts, so the skill does not run local programs of its own.
  • Network: yes, it uses web search and an external reviewer through the bridge.
  • Keep a human literature review before submitting any paper.

What people use it for

  • Deciding whether a method idea is worth a pilot experiment before months of work.
  • Getting a structured first pass on prior work before writing a research proposal.
  • Checking whether a claimed contribution is already in a named recent paper.
  • Having a Claude model act as an independent reviewer for a Codex-based workflow.

Frequently asked questions

Is the Novelty Check skill free to use?

The source lists the licence as MIT, which is a permissive open-source licence. The source does not state any price, so check the repository for any terms beyond the licence itself.

What do I need to run the Novelty Check skill?

You need Claude Code or a compatible agent with WebSearch and WebFetch tools, plus the claude-review MCP bridge running as a local MCP server. CLAUDE_REVIEW_MODEL and oracle-pro are optional.

Where do I install the novelty-check skill?

Unzip the downloaded package into ~/.claude/skills/ for personal use or into .claude/skills/ in a project. In the Claude apps, upload it under Settings, then Capabilities, then Skills.

Can the novelty check replace a literature review?

No. The skill describes itself as an early gate and says it can miss papers or misjudge them. A human literature review is still needed before publication.

Why does the skill need an external reviewer?

The skill sends the idea and the list of papers to a separate Claude reviewer through the claude-review bridge, so the judgement comes from a different reviewer than the one doing the search. The source describes this as cross-family review when Codex is the executor.

Why did the skill give my idea PROCEED WITH CAUTION instead of ABANDON?

The skill only recommends ABANDON when a specific published paper already contains the result, and it must name that paper. Crowded ground with a stated difference is treated as PROCEED, and PROCEED WITH CAUTION is used only when the difference is thin or contested.

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 this skill if you are a researcher, student or Codex user who wants a structured, traceable first pass on prior work before a pilot experiment or proposal. Skip it if your project cannot send queries or review content to external services, if you need a final publication-grade novelty verdict, or if your task is not about academic novelty of methods.

Download the skill