AI Research Skills
A 98-skill open-source library (as of July 2026) that turns Claude Code (or Codex, Gemini CLI, and other coding agents) into an AI/ML research agent — covering fine-tuning, distributed training, inference serving, evaluation, RAG, and paper writing, orchestrated end to end by an autoresearch layer.
Last reviewed by the paperbanana team on Jul 13, 2026
Install
npx @orchestra-research/ai-research-skillsWhat is AI Research Skills?
AI Research Skills (published as the AI-Research-SKILLs repo by Orchestra Research) is an open-source library of 98 skills (as of July 2026) organized into 23 categories that span the full AI/ML research lifecycle — from ideation and literature survey through fine-tuning, distributed training, evaluation, and inference, all the way to writing the resulting paper. Unlike general academic-writing assistants, its focus is squarely on AI research engineering: it packages deep, production-grade guidance for frameworks researchers actually use day to day, such as Megatron-Core, vLLM, TRL/GRPO, Hugging Face Tokenizers, DeepSpeed, and Axolotl.
Each skill follows a standardized structure: a SKILL.md quick reference (50–150 lines covering when to use the skill and quick patterns), a references/ folder with 300KB+ of deep documentation sourced from official repos and real GitHub issues, and optional scripts/ and assets/ folders. According to the repo’s stated quality standards, skills document version history and breaking changes and link back to official docs rather than paraphrasing from memory, and the roadmap tracks metrics like average lines per skill and how many skills reach "gold standard" reference depth.
At the center of the library sits the Autoresearch skill, a central orchestration layer that uses a two-loop architecture (an inner optimization loop and an outer synthesis loop) to manage a project end to end — literature survey, ideation, experiments, synthesis, and paper writing — while automatically routing to whichever of the other 97 domain skills a given step needs. The README documents two demo papers produced this way, including one where the orchestrated agent refuted its own initial hypothesis mid-run and pivoted to a stronger finding.
Core capabilities
98 skills across 23 categories
Coverage spans Model Architecture, Fine-Tuning, Mechanistic Interpretability, Distributed Training, Optimization, Evaluation, Inference & Serving, MLOps, Agents, RAG, Multimodal, Prompt Engineering, Safety & Alignment, Ideation, and ML Paper Writing — the full pipeline from idea to trained, served, and evaluated model.
Autoresearch orchestration layer
A dedicated Autoresearch skill runs a two-loop architecture (inner optimization + outer synthesis) that manages the full research lifecycle and routes automatically to the right domain skill at each stage, supporting continuous operation via Claude Code /loop or a cron-style heartbeat.
Standardized, deep-reference skill format
Every skill ships a concise SKILL.md plus a references/ directory (300KB+ target) with API docs, tutorials, real GitHub issues and fixes, release/breaking-change history, and codebase navigation notes — aimed at production-ready depth rather than shallow prompts.
ML Paper Writing and Academic Plotting
A dedicated ML Paper Writing skill provides LaTeX templates for NeurIPS, ICML, ICLR, ACL, AAAI, and COLM plus a citation-verification workflow, paired with an Academic Plotting skill that generates architecture diagrams (via Gemini AI) and data-driven charts (matplotlib/seaborn) in venue-specific styles.
Agent-Native Research Artifact (ARA) category
Three skills — ARA Compiler, ARA Research Manager, and ARA Rigor Reviewer — compile papers, repos, or experiment logs into a structured, falsifiable research artifact with claims, an exploration graph, and evidence, then score it against six dimensions of research rigor.
One-command, multi-agent install
The npx @orchestra-research/ai-research-skills installer auto-detects installed coding agents (Claude Code, OpenCode, Cursor, Codex, Gemini CLI, Qwen Code, and more), installs skills to ~/.orchestra/skills/ with per-agent symlinks, and supports listing, updating, and selective uninstall.
What you can use it for
Fine-tuning a model on custom data
The README’s own example: "I need to fine-tune Llama 3 with custom data" routes to the Axolotl skill for YAML-based configs across 100+ supported models.
Optimizing inference latency
"How do I optimize inference latency?" routes to the vLLM skill, which documents PagedAttention and batching for high-throughput serving.
Scaling training across many GPUs
"We need to scale training to 100 GPUs" routes to the DeepSpeed skill, covering ZeRO stages and 3D parallelism for distributed training teams.
Running an end-to-end autonomous research project
Load the Autoresearch skill to drive literature survey, ideation, experiments, and paper writing in one continuous loop — the repo’s own demos used this path to produce two full papers, including one where the agent pivoted after refuting its initial hypothesis.
Interpreting what a trained model actually learned
Reach for the Mechanistic Interpretability category (TransformerLens, SAELens, pyvene, nnsight) to run activation caching, sparse-autoencoder feature discovery, or causal interventions on model internals.
How to get started
- 1
Run the installer
Run npx @orchestra-research/ai-research-skills for an interactive installer that auto-detects your coding agent(s) and offers everything, a quickstart bundle, by-category, or individual-skill installation.
- 2
Or hand an agent the welcome doc
For a fully agent-driven setup, tell your AI agent to "Read https://www.orchestra-research.com/ai-research-skills/welcome.md and follow the instructions" — this installs all 98 skills and loads the Autoresearch orchestration layer directly.
- 3
Install by category via the Claude Code marketplace (alternative)
Run /plugin marketplace add orchestra-research/AI-research-SKILLs, then install a category, e.g. /plugin install fine-tuning@ai-research-skills or /plugin install inference-serving@ai-research-skills, if you only need specific parts of the library.
- 4
Let Autoresearch route the work, or update later
Once installed, describe your research goal and Autoresearch routes to the right domain skill automatically; run npx @orchestra-research/ai-research-skills update whenever you want the latest skill versions.
How it compares to similar skills
AI Research Skills is built for AI/ML research engineering, not general academic writing — its 98 skills center on training, evaluation, and inference frameworks. For adjacent needs, these other entries in the directory may fit better.
Academic Research Skills
Pick Academic Research Skills if your bottleneck is the writing pipeline itself (drafting, peer-review-style critique, citation integrity gates) rather than ML training/evaluation tooling.
Scientific Agent Skills
Pick Scientific Agent Skills if you work in wet-lab or data-heavy natural sciences and need domain research tools outside the ML/engineering stack this library targets.
LaTeX Document Skill
Pick latex-document-skill if you specifically need dedicated LaTeX typesetting/formatting help rather than the broader research-engineering skill set bundled here.