🧬 EvoSkills

The official skill repository for EvoScientist. Each skill is an installable knowledge pack that extends EvoScientist with domain-specific expertise.

📦 Installation

[!IMPORTANT] These skills are purpose-built for EvoScientist — together they amplify each other, unlocking the full potential of both the agent and the skills. Under EvoScientist, skills evolve across research cycles through persistent memory (evo-memory).

In-session commands

Install all skills at once:

/install-skill EvoScientist/EvoSkills@skills

Or install a single skill:

/install-skill EvoScientist/EvoSkills@skills/paper-planning

Ask EvoScientist directly

Simply ask the agent in conversation:

"Install all skills from EvoScientist/EvoSkills@skills."

[!TIP] Not using EvoScientist? These skills are compatible with any coding agent. One command via skills.sh to install on Claude Code, OpenCode, Cursor, Codex, Gemini CLI, DeepAgents, and more:

npx skills add EvoScientist/EvoSkills

✨ Available Skills

Paper Suite + Self-Evolution Suite: Each skill is self-contained — use them individually or combine freely. The self-evolution loop now runs through research-ideation, experiment-pipeline, and evo-memory.

🔌 MCP Server Marketplace

The mcp/ directory contains a curated collection of MCP servers that extend agents with external tools — web search, academic paper retrieval, documentation lookup, and more. Browse the full list or install directly:

/install-mcp              # interactive browser
EvoSci mcp install arxiv  # install by name

⛳️ Framework Overview

The diagram above shows the full EvoScientist pipeline. The Researcher Agent (top, blue) runs idea tree search and Elo tournament ranking to produce a research proposal. The Engineer Agent (bottom, green) executes the 4-stage experiment pipeline. The Evolution Manager Agent (right) manages three memory evolution mechanisms — IDE, IVE, and ESE — that feed learned knowledge back into Ideation Memory (M_I) and Experimentation Memory (M_E) for future cycles.

🎢 Skill Pipeline

flowchart LR
    A["<b>🔬 Research Phase</b><br/>💡 research-ideation"]
    --> B["<b>⚙️ Experiment Phase</b><br/>📐 paper-planning<br/>🧪 experiment-pipeline<br/>🔧 experiment-craft<br/>🔄 experiment-iterative-coder"]
    --> C["<b>📝 Writing Phase</b><br/>✍️ paper-writing<br/>🔍 paper-review<br/>💬 paper-rebuttal<br/>🎤 academic-slides"]

    D[("🧠 evo-memory<br/>(IDE · IVE · ESE)")] <--> A
    D <--> B
    E["📚 paper-navigator<br/>(standalone)"] -.-> A
    E -.-> B
    F["🍌 nano-banana<br/>(standalone)"] -.-> C
    G["🧮 evomath-tao<br/>(standalone)"] -.-> A
    G -.-> B

    style A fill:#7C3AED,stroke:#5B21B6,stroke-width:2px,color:#fff
    style B fill:#D97706,stroke:#B45309,stroke-width:2px,color:#fff
    style C fill:#16A34A,stroke:#15803D,stroke-width:2px,color:#fff
    style D fill:#475569,stroke:#334155,stroke-width:2px,color:#fff
    style E fill:#0369A1,stroke:#075985,stroke-width:2px,color:#fff
    style F fill:#D97706,stroke:#B45309,stroke-width:2px,color:#fff
    style G fill:#BE185D,stroke:#9D174D,stroke-width:2px,color:#fff

💡 research-ideation — Literature Grounding, Tournament & Proposal

The starting point of the research pipeline. It now covers the full path from literature grounding to ranked ideas to a concrete proposal:

• Load Prior Knowledge — Read evo-memory first to reuse feasible directions and avoid known dead ends
• Literature Grounding — Use paper-navigator to collect and analyze papers before generating ideas
• Multi-Track Ideation + Refinement — Generate candidates across multiple personas, then iteratively strengthen them
• Elo Tournament — Rank refined ideas on novelty, feasibility, relevance, and clarity; present the top-3
• Proposal Extension — Expand the selected winner into a manuscript-quality research proposal

📝 research-survey — Literature Survey & Synthesis

Dedicated skill for turning a large paper collection into a structured survey report:

• Adaptive Outline — Generate a field-specific outline based on the query type and literature set
• Draft + Expansion Pipeline — Draft from top papers, then deepen each section with the full collection
• Summary Refinement — Build section summaries before rewriting the abstract, introduction, and conclusion
• Survey-Grade Output — Comparative tables, taxonomy-based method organization, dense citations, and references

📐 paper-planning — Research Paper Planning & Outline Generation

Guides pre-writing planning before a single word is drafted. Covers four key activities:

• Story Design — Reverse-engineer the narrative: task → challenge → insight → contribution → advantage
• Experiment Planning — Plan comparisons, ablations, and demo scenarios with structured checklists
• Figure Design — Pipeline figures that highlight novelty; teaser figures that hook reviewers
• Timeline Management — 4-week countdown schedule from outline to submission

Includes counterintuitive tactics: write your rejection letter first, narrow claims before broadening, and plan fallback narratives.

🧪 experiment-pipeline — 4-Stage Experiment Execution

A structured framework for executing research experiments with attempt budgets and gate conditions:

• Stage 1: Initial Implementation — Get baseline code running and reproduce known results (≤20 attempts)
• Stage 2: Hyperparameter Tuning — Optimize configuration for your setup (≤12 attempts)
• Stage 3: Proposed Method — Implement and validate the novel method (≤12 attempts)
• Stage 4: Ablation Study — Prove each component's contribution (≤18 attempts)
• Code Trajectory Logging — Structured attempt logging that feeds into evo-memory
• Counterintuitive Rules — Initial implementation is not wasted time; budget limits prevent rabbit holes; failed attempts are data

Integrates with experiment-craft for failure diagnosis within stages and evo-memory for cross-cycle learning.

🔧 experiment-craft — Experiment Debugging & Iteration

A systematic approach to experiment debugging, logging, and iterative improvement:

• 5-Step Diagnostic Flow — Collect failures → find a working version → bridge the gap → hypothesize → fix
• Counterintuitive Rules — Change one variable at a time; effective experiments beat more experiments
• Experiment Logging — 5-section structured log template for reproducible records
• Handoff to Paper-Writing — Feed validated results and logs into paper-writing for drafting

🔄 experiment-iterative-coder — Iterative Code Refinement

Structured plan → code → evaluate → refine cycles for higher code quality:

• Phase Decomposition — Break complex tasks into 1-5 sequential phases
• Iteration Loop — Up to 3 iterations per phase (10 total): plan, code, run lint/tests, score, decide
• Objective Evaluation — ruff lint + pytest with dynamic score weighting and hard caps
• Failure Mode Guidance — Targeted responses for timeout, syntax, import, test, and lint failures

Integrates with experiment-craft for stuck diagnoses and evo-memory for loading prior strategies.

✍️ paper-writing — Section-by-Section Paper Drafting

A proven 11-step workflow for writing academic papers with LaTeX templates:

• Structured Process — From pipeline sketch → story design → Method → Experiments → Related Work → Abstra