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 a set of instructions that tells an AI agent how to test whether a research idea has already been done. You describe a method, and the agent breaks it into core claims, searches the literature, reads the abstracts of likely matches and writes a structured report with the closest prior papers and a recommendation. This guide covers what the skill does, how to install it, how to trigger it, and what it needs and exposes before you rely on it.
Novelty Checkwanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download the ZIP. Download the novelty-check ZIP from this page and unzip it on your computer.
- 2Place the folder. Move the novelty-check folder into ~/.claude/skills/ for all projects, or into a project's .claude/skills/ folder for one project.
- 3Or upload in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the skill.
- 4Describe your idea. Write the method in a few sentences, including the mechanism and what it is meant to improve over baselines.
- 5Ask for a novelty check. Use a trigger phrase such as "novelty check" or "has anyone done this", or the Chinese phrase 查新.
- 6Review the report. Read the closest prior work table and the stated delta, and open the cited papers yourself.
- 7Decide before building. Use the recommendation as one input to whether you run a pilot, and keep other checks such as external review in the loop.
What the novelty-check skill makes the agent do
An agent skill is a folder with a SKILL.md file that gives an agent a repeatable procedure for a task. When this skill is active and you describe a research idea, the agent first extracts three to five core technical claims: the method, the problem it solves, the mechanism, and what separates it from obvious baselines.
For each claim, the agent runs web searches across arXiv, Google Scholar and Semantic Scholar, checks recent conference papers, and fetches the abstracts of likely overlaps. It can also send a written dossier to a second model through a reviewer tool for an independent check.
- Core-claim extraction from your description
- Multi-source search with several query formulations per claim
- Abstract reading for likely matches
- A report with closest prior papers, the delta from each, a novelty score and a recommendation to proceed, proceed with caution, or abandon
Install the novelty-check skill
Download the ZIP from this page and unzip it. To make the skill available across all your projects, place the novelty-check folder inside ~/.claude/skills/. To limit it to one project, place it inside that project's .claude/skills/ folder.
If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/novelty-check/
- Project install: .claude/skills/novelty-check/ inside your project
Trigger the skill with natural requests
The skill's description lists phrases that activate it. You can write in English or Chinese. Describe the idea in enough detail that the agent can pull out its core mechanism; a one-line topic gives weaker searches.
The skill also accepts an optional argument hint for the method or idea description.
- "Novelty check: I want to use a learned gate to skip layers during decoding."
- "Has anyone done this? I plan to fine-tune with synthetic critiques for code review."
- "查新: 用扩散模型生成表格数据的缺失值"
- "Check novelty of my idea before I start a pilot experiment."
What the skill needs
The skill's allowed tools include Bash, WebSearch, WebFetch, Grep, Read and Glob. Web search and fetch are required for the literature sweep. Shell and file tools are used to read and write files such as the dossier.
The reviewer step is optional. It uses the Codex MCP tool (mcp__codex__codex) or another configured reviewer server, and the skill names a model called gpt-6-astra, which requires an OpenAI model account on the Codex route. The skill also references a shared reviewer-backends file and a reviewer_route.py helper, which are not included in the files provided with this skill. Without them, the reviewer step may not run as written.
- Web search and fetch tools (WebSearch, WebFetch)
- Shell and file tools (Bash, Read, Grep, Glob)
- Optional: Codex MCP or another reviewer MCP server for cross-model checks
- Optional: an OpenAI account if you use the Codex reviewer route
Safety notes: network use and what gets sent
The skill uses the network. It sends your search queries to public web search services and fetches pages. If the reviewer step runs, it writes a dossier containing your method description, core claims and the candidate papers, and sends that to an external reviewer model.
Do not use this skill for unpublished or confidential ideas that must stay private. The skill does not run bundled scripts, but it does allow Bash, so review what the agent executes if you are working in a sensitive environment.
- Search queries go to the web
- A method dossier may go to an external reviewer model
- Do not use for confidential or unpublished work you must keep private
Read the report and treat it as an early gate
The report lists the core claims, a table of closest prior work with overlap and key differences, a novelty score out of 10, and a recommendation. The skill's own rules say the check is an early gate, not final clearance. Search coverage is limited to what the agent can find online, so relevant papers can be missed.
Papers in the prior-work table are meant to be verified before they appear. If verification is unavailable, the skill says it will tag entries as UNVERIFIED rather than drop them, so check each citation yourself before relying on it.
- PROCEED: the idea has neighbors but a stated delta a reviewer could verify
- PROCEED WITH CAUTION: the delta is thin or contested, and the report says what would strengthen it
- ABANDON: the skill requires naming a specific published paper that already contains the result
What people use it for
- Deciding whether a method idea is worth a pilot experiment before committing months of work
- Checking whether a student project topic has already been published
- Getting a quick literature sweep with a written summary of the closest prior papers
- Stress-testing a claimed contribution by asking for the specific delta from each neighbor
Frequently asked questions
How do I install the novelty-check skill?
Download the ZIP from this page and unzip it. Put the novelty-check folder in ~/.claude/skills/ for all projects, or in a project's .claude/skills/ folder for one project. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
Is the novelty-check skill free?
The source material provided for this page lists the skill's licence as MIT, but it does not give pricing details. The optional Codex reviewer route requires an OpenAI model account, which is a separate arrangement.
What does the novelty-check skill need to run?
It needs web search and fetch tools (WebSearch, WebFetch) and shell and file tools (Bash, Read, Grep, Glob). The optional cross-model step needs the Codex MCP tool or another reviewer server, and an OpenAI account if you use the Codex route.
Does the skill send my research idea anywhere?
It sends search queries to the web. If the reviewer step runs, it may write a dossier with your method description and send it to an external reviewer model. Avoid it for confidential or unpublished ideas that must stay private.
Can I rely on the skill to confirm an idea is novel?
No. The skill describes itself as an early gate, not final novelty clearance, and its search is limited to what the agent can find online. Missed papers are possible, so check the cited work yourself and consult others before committing to a project.
Does the skill work with Chinese queries?
Yes. The trigger phrases include 查新 and 有没有人做过, so you can ask in Chinese as well as English.
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
This skill suits researchers and students who want a fast, structured literature sweep and a written list of closest prior work before starting a pilot. It is not suitable as final novelty clearance, for general web research, or for confidential ideas you must not send to outside search or reviewer services. If you need authoritative novelty clearance, pair it with your own database searches and expert review.
Download the skill