UditAkhourii/neuroarxiv

A skill to kill from-scratch coding — Claude checks real arXiv prior art before it designs a new architecture.

What it solves

NeuroArxiv prevents developers from "building from scratch" by forcing an AI agent to find and commit to a specific, grounded recommendation based on existing prior art from arXiv. Unlike general search tools that provide a list of sources for the user to filter, NeuroArxiv uses a disciplined process to synthesize multiple papers into a single, cited architectural or algorithmic decision, specifically flagging documented weaknesses and limitations of the cited research to avoid costly rebuilds.

How it works

The system follows a strict "isolate-then-converge" pipeline:

  1. Categorize: Maps the user's problem to specific arXiv categories and search terms.
  2. Fetch: Performs deterministic HTTP requests to the arXiv export API to gather papers.
  3. Diverge: An LLM reads each paper's abstract in complete isolation from the others to prevent source anchoring.
  4. Score & Cluster: Papers are scored for relevance and rigor, then grouped by architectural approach.
  5. Converge: The system selects one recommended path, synthesizes the findings, provides a first step for implementation, and lists pitfalls derived from all analyzed papers.

Who it’s for

It is designed for developers and researchers committing to non-trivial architectures, algorithms, or systems techniques where the cost of a wrong guess is a full rebuild.

Highlights

  • Decision-Driven: Moves beyond search by forcing a single, grounded recommendation rather than a list of links.
  • Source Skepticism: Specifically identifies and flags limitations and withdrawn proofs within the cited sources.
  • Agent Integration: Installs as a skill for Claude Code and Codex CLI.
  • Grounded Metadata: Uses real arXiv API metadata to ensure citations and links are authentic.

Related

  • Project
    juliye2025/evil-read-arxiv

    An automated research workflow that searches, recommends, and analyzes academic papers from arXiv and Semantic Scholar, integrating the results directly into Obsidian notes.

  • Project
    DeepXiv/deepxiv_sdk

    A CLI and Python SDK for agentic search over scientific literature, providing structured, pre-parsed paper sections and cited answers to reduce token costs and research friction.

  • Project
    Nutlope/open-deep-research

    An AI-powered tool that generates comprehensive, sourced research reports by iteratively searching the web and synthesizing information.

  • Project
    FSoft-AI4Code/CodeWiki

    CodeWiki is an open‑source Python tool that uses large language models to automatically generate comprehensive, architecture‑aware documentation (markdown + Mermaid diagrams) for whole codebases across ten languages. It hierarchically decomposes a repository, runs a recursive multi‑agent LLM pipeline, and outputs ready‑to‑publish docs, with support for many LLM providers and a rich CLI.

  • Project
    federicodeponte/opendraft

    An open-source AI research engine that generates long-form, source-grounded academic drafts with citations verified against CrossRef, OpenAlex, and Semantic Scholar.