EvoScientist/EvoSkills

What it solves

EvoSkills provides a comprehensive library of installable "knowledge packs" designed to automate and enhance the academic research lifecycle. It solves the problem of fragmented research workflows by providing structured, agentic protocols for everything from initial literature discovery and idea generation to experimental execution, paper writing, and peer-review rebuttals.

How it works

Each skill is a self-contained module that can be installed into an AI agent (specifically optimized for EvoScientist, but compatible with other coding agents via skills.sh). These skills implement rigorous, multi-stage workflows:

• Research & Ideation: Uses literature grounding and Elo tournaments to rank and refine research proposals.
• Experimentation: Implements a 4-stage execution pipeline (Implementation, Tuning, Proposed Method, Ablation) with attempt budgets and structured logging.
• Writing & Review: Provides 11-step drafting workflows, LaTeX templates, and adversarial self-review strategies to simulate rejection and improve quality.
• Memory & Evolution: The evo-memory skill creates a persistent layer that tracks successful and failed research directions, allowing the agent to evolve its strategies across multiple research cycles.
• Specialized Tools: Includes dedicated modules for generating publication-ready Matplotlib figures, creating academic slides, and a Tao-style rigorous proof workflow for mathematics.

Who it’s for

Academic researchers, AI scientists, and developers building autonomous research agents who need a standardized, high-quality framework for conducting scientific inquiry and drafting publications.

Highlights

• End-to-End Pipeline: Covers the entire research loop: ideation $\rightarrow$ experimentation $\rightarrow$ writing $\rightarrow$ review.
• Self-Evolution: Persistent memory mechanisms (IDE, IVE, ESE) that allow the agent to learn from failures and successes across cycles.
• Rigorous Protocols: Uses "counterintuitive" rules and structured budgets to prevent agents from falling into rabbit holes during coding or experimentation.
• Broad Tooling: Includes MCP servers for external tool integration (e.g., arXiv, web search) and specialized visual generation for presentations.
• Mathematical Rigor: A dedicated Olympiad-grade proof workflow with calibrated abstention to avoid "hallucinated" proofs.