Orchestra-Research/AI-Research-SKILLs

Metadata

• **Stars**: 6,637
• **License**: MIT
• **Last pushed**: 2026-04-10
• **Topics**: ai, ai-research, claude, claude-code, claude-skills, codex, gemini, skills, vllm
• **Fork**: No

Overview

Comprehensive open-source library of 87 AI research skills organized into 22 categories covering the full AI research lifecycle — from idea to paper. Includes an autoresearch orchestration skill that uses a two-loop architecture (inner experiment loop + outer synthesis loop) to run autonomous research projects.

Architecture

• 22 numbered skill categories (model-architecture, tokenization, fine-tuning, interpretability, data-processing, post-training, safety, distributed-training, infrastructure, optimization, evaluation, inference, mlops, agents, rag, prompt-engineering, observability, multimodal, emerging-techniques, paper-writing, ideation)
• Each skill follows SKILL.md format with YAML frontmatter, references/, scripts/, templates/
• Claude Code plugin marketplace with marketplace.json
• NPM installer: npx @orchestra-research/ai-research-skills

Extractable Value

Process: specializations/science/ai-research-orchestration

The autoresearch SKILL.md contains a complete two-loop research orchestration methodology: 1. Bootstrap: scope question, literature survey, form hypotheses 2. Inner loop: pick hypothesis, experiment, measure, record, learn 3. Outer loop: review results, find patterns, update findings, new hypotheses 4. Finalize: write paper, presentation, archive

Includes workspace structure template (research-state.yaml, experiments/, paper/), agent continuity patterns, and routing to domain skills. This is a full generic research methodology suitable for specializations/science/autonomous-ai-research.

Process: specializations/science/ml-paper-writing

The research-paper-writing skill (also found in hermes-agent, attributed to Orchestra Research) provides a complete 7-phase pipeline for ML paper writing targeting NeurIPS/ICML/ICLR/ACL/AAAI/COLM. Includes experiment design, monitoring, statistical analysis, iterative drafting, self-review, and submission prep.

Plugin idea: ai-research-skills

The marketplace already has a marketplace.json with plugin structure. Could be assimilated as a babysitter marketplace plugin providing AI research domain skills. However, these are tool/framework-specific skills (vLLM, DeepSpeed, etc.), not babysitter processes — they'd need to be wrapped as process definitions.

Processes

1. Autonomous AI Research Orchestration

• **Source**: autoresearch SKILL.md two-loop architecture (inner experiment loop + outer synthesis loop)
• **Placement**: specializations/science/autonomous-ai-research.js
• **Description**: Complete research methodology: bootstrap (scope question, literature survey, form hypotheses) → inner loop (pick hypothesis, experiment, measure, record, learn) → outer loop (review results, find patterns, update findings, new hypotheses) → finalize (write paper, presentation, archive).

2. ML Paper Writing Pipeline

• **Source**: research-paper-writing skill 7-phase pipeline for ML conferences
• **Placement**: specializations/science/ml-paper-writing.js
• **Description**: End-to-end ML paper writing for NeurIPS/ICML/ICLR/ACL/AAAI/COLM: experiment design → monitoring → statistical analysis → iterative drafting → self-review → submission prep.

Plugin Ideas

• **AI Research Skills Suite**: Babysitter marketplace plugin providing AI research domain skills with 87 skills across 22 categories covering full AI research lifecycle.

Library Mapping

Plugin Marketplace Mapping

Classification

• **Archetype**: Domain skill library with orchestration layer
• **Primary value**: Two processes extractable — autonomous research orchestration and ML paper writing pipeline
• **Process placement**: specializations/science/autonomous-ai-research and specializations/science/ml-paper-writing