Give your coding agent guidance for AI research, from planning experiments and training models to evaluating results and writing papers.
AI Research Skills supplies coding agents with instructions, examples and reference material for research and engineering tasks. It covers model development, training, evaluation, deployment and research writing. Installed domain skills can activate when you discuss relevant topics with your agent.
For a project that spans several stages, explicitly ask your agent to load autoresearch. That skill coordinates experiment cycles, synthesizes findings and routes work to domain skills. Installing the library makes its guidance available; running experiments still depends on the software, data and compute your task needs.
Load autoresearch with a research question or an existing plan. Its instructions guide the agent through literature search, hypothesis testing and synthesis, with research state and findings kept between sessions. Continuous operation uses a separate agent loop or scheduler.
Use the Axolotl, LLaMA-Factory, PEFT or Unsloth guidance to prepare a fine-tuning configuration. For instruction tuning, preference alignment or reward-based training, the TRL and reinforcement learning skills cover the corresponding training workflows.
When a model exceeds one GPU's memory or training needs several machines, the distributed training skills cover sharding, parallelism and checkpointing with tools such as FSDP2, DeepSpeed and Megatron-Core.
Use the evaluation skills to set up language-model or code-generation benchmarks. The serving and quantization skills then help you explore deployment choices for your hardware, including vLLM, TensorRT-LLM and llama.cpp.
Use TransformerLens or nnsight to inspect model internals, pyvene to test causal interventions, and SAELens to study features in model activations.
Turn research results into an ML or systems paper with guidance on argument structure, citations and LaTeX. Academic Plotting covers figures; Presenting Conference Talks builds slides, speaker notes and a talk script from a compiled paper.
Use Claude Code's agent-native-research-artifact category plugin for this workflow. Compile papers, code and experiment notes into an Agent-Native Research Artifact. Record session decisions with the research manager, then use the rigor reviewer after structural validation to assess the artifact's claims and supporting evidence.
These skills give your coding agent instructions and examples for the tasks below. The model libraries, datasets, compute and service credentials needed to carry out a task are supplied separately. In Claude Code's marketplace, the skills are packaged into category plugins; each category is installed as a whole plugin.
Showing all 96
Implement, train and fine-tune language models with LitGPT.
Explore selective state-space models and Mamba architectures.
Learn GPT architecture through a compact implementation and training examples.
Work with RWKV's recurrent architecture for language modeling.
Configure distributed LLM pretraining with TorchTitan and multidimensional parallelism.
Train custom tokenizers and configure text encoding with Hugging Face Tokenizers.
Train multilingual BPE or Unigram tokenizers directly from raw text.
Configure LLM fine-tuning with Axolotl, including LoRA and QLoRA.
Fine-tune language and multimodal models through LLaMA-Factory's interface.
Adapt large models with LoRA, QLoRA and other parameter-efficient methods.
Configure Unsloth fine-tuning with LoRA and QLoRA.
Inspect and manipulate model internals, with optional remote execution through NDIF.
Test causal hypotheses with activation patching and other model interventions.
Train and analyze sparse autoencoders to study features in model activations.
Inspect transformer activations, attention patterns and circuits with TransformerLens.
Clean, filter and deduplicate training datasets with NeMo Curator.
Build distributed preprocessing, data-loading and batch-inference pipelines with Ray Data.
Train reasoning and task-specific models with GRPO in TRL.
Configure large-scale reinforcement learning with miles, including quantized MoE training.
Set up distributed RLHF and preference training with OpenRLHF.
Align language models with reference-free Simple Preference Optimization.
Combine Megatron and SGLang for reinforcement learning with slime.
Build PyTorch reinforcement learning experiments with torchforge's infrastructure and algorithms.
Use TRL for instruction tuning, preference alignment and reward-based training.
Configure scalable LLM reinforcement learning with verl.
Apply self-critique, revision and AI feedback to safety alignment.
Classify model inputs and outputs for safety moderation with LlamaGuard.
Define runtime input, output and conversational safeguards with NeMo Guardrails.
Detect prompt injection and jailbreak attempts in prompts and retrieved content.
Add distributed execution and mixed precision to PyTorch training scripts.
Configure ZeRO optimization, pipeline parallelism and distributed training with DeepSpeed.
Train large language models with Megatron-Core's parallelism strategies.
Shard PyTorch models with FSDP2 and configure distributed checkpoints.
Structure training with Lightning's Trainer, callbacks and distributed execution.
Run distributed model training and hyperparameter sweeps across machines.
Provision dedicated GPU instances and multi-node clusters on Lambda.
Run GPU jobs, batch workloads and model APIs on Modal.
Run training across clouds with spot recovery and GPU cost optimization.
Compress model weights with activation-aware quantization for constrained GPU memory.
Use low-bit model loading, QLoRA and memory-saving optimizers with bitsandbytes.
Configure memory-efficient attention with Flash Attention or PyTorch SDPA.
Prepare quantized GGUF models for llama.cpp on consumer hardware.
Apply post-training weight quantization to reduce LLM memory requirements.
Quantize language models with HQQ without a calibration dataset.
Choose training recipes and debug loss spikes, memory issues and throughput.
Benchmark code generation with HumanEval, MBPP and multilingual evaluation tasks.
Compare language models on academic benchmarks with lm-evaluation-harness.
Configure container-based model evaluation on local, cluster or cloud infrastructure.
Run local LLM inference on CPUs, Apple Silicon and supported GPUs.
Serve language models with prefix caching and constrained generation.
Optimize and serve language models on NVIDIA GPUs with TensorRT-LLM.
Serve LLM APIs with continuous batching and tensor parallelism in vLLM.
Track experiments, version registered models and manage model deployment with MLflow.
Visualize training metrics, inspect model graphs and profile performance.
Track experiments, run hyperparameter sweeps and manage model artifacts with W&B.
Build and deploy persistent agents and visual automation workflows with AutoGPT.
Coordinate role-based agent teams in sequential or hierarchical workflows.
Build applications with model calls, retrieval, memory and tool-using agents.
Build document ingestion, indexing and question-answering pipelines with LlamaIndex.
Store embeddings and metadata for document retrieval with Chroma.
Index dense vectors for similarity search and clustering with FAISS.
Build managed vector retrieval with hybrid search, metadata filters and namespaces.
Configure vector retrieval and filtered hybrid search with Qdrant.
Generate text and image embeddings for retrieval, similarity and clustering.
Build modular language-model pipelines and optimize their prompts with DSPy.
Constrain model generation with grammars, regular expressions and structured formats.
Extract structured data with Pydantic validation, retries and partial-result streaming.
Define structured model outputs with schemas and constrained generation.
Trace application runs and evaluate model outputs against datasets with LangSmith.
Inspect LLM traces and evaluate application behavior with Phoenix.
Generate music and sound effects from text with AudioCraft.
Use BLIP-2 for image captioning, retrieval and visual question answering.
Match images and text for classification and cross-modal retrieval.
Evaluate Cosmos Policy robot manipulation in LIBERO and RoboCasa simulations.
Build image-based conversations and visual question answering with LLaVA.
Fine-tune OpenPI robot policies and serve them for action inference.
Fine-tune and evaluate OpenVLA-OFT robot action policies.
Segment image objects using point, box or mask prompts.
Generate, transform and inpaint images with Stable Diffusion through Diffusers.
Transcribe speech, identify spoken languages and translate audio into English.
Transfer model behavior from a larger teacher to a smaller student.
Extend model context windows with positional encoding and interpolation techniques.
Combine fine-tuned model capabilities with mergekit without retraining.
Reduce model size with structured or unstructured weight pruning.
Configure mixture-of-experts models, routing, load balancing and expert parallelism.
Explore draft-model and parallel decoding methods to reduce inference latency.
Coordinate research experiments, synthesize findings and route work to domain skills.
Structure ML papers, verify citations and prepare LaTeX submissions.
Create research diagrams and data-driven figures for conference papers.
Plan and write systems papers with venue-specific structures and checklists.
Create conference slides, speaker notes and a script from a compiled paper.
Explore research directions through structured brainstorming frameworks.
Generate research ideas through analogy, recombination and changes to constraints.
Compile research inputs into structured claims, evidence and an exploration graph.
Record decisions and experiments with provenance tags at a session's end.
Review an artifact's research rigor after its structural validation passes.
No matching results.
In Claude Code, add the marketplace and install its category plugins. Alternatively, use the Orchestra installer for your coding tool. The installer omits the compiler, research-manager and rigor-reviewer skills; those are available in Claude Code's agent-native-research-artifact category plugin.
Use this method in a local Claude Code session. Adding the marketplace makes its category plugins available to install; it does not install their skills. Install the categories you need, or repeat the installation for every category to get the full library. Each category plugin installs as a whole.
Run this in your Claude Code session to add the AI Research Skills marketplace.
/plugin marketplace add orchestra-research/AI-research-SKILLsChoose a category plugin by its name:
model-architecturetokenizationfine-tuningmechanistic-interpretabilitydata-processingpost-trainingsafety-alignmentdistributed-traininginfrastructureoptimizationevaluationinference-servingmlopsagentsragprompt-engineeringobservabilitymultimodalemerging-techniquesautoresearchml-paper-writingideationagent-native-research-artifactRun this in the same Claude Code session. For another category, replace distributed-training with its name from the list above and keep @ai-research-skills unchanged. The example selects the distributed-training plugin, not the full library.
/plugin install distributed-training@ai-research-skillsIn the plugin details panel, review the plugin, choose user, project or local scope, and complete the installation. Repeat for any additional category plugins you want.
Read the installation summary. If it says the plugin is active now, no reload is needed. If it says a reload is needed, close the panel to let Claude Code apply it. If Claude Code leaves the plugin pending after a cache warning, start a new session before checking availability.
Confirm that each category plugin you installed appears in the plugin panel's Installed tab. In the session, type / and look for that plugin's skills under its name. Locally installed plugins are not loaded by Claude Code cloud sessions.
Orchestra marketplace instructions · Category plugin manifest · Claude Code plugin installation
Use a computer with Claude Code configured, Node.js 18 or later, Git and access to GitHub. The installer detects Claude Code through its existing ~/.claude directory. Every scope below installs skill files from categories 0 and 01–21, including autoresearch. It omits the compiler, research-manager and rigor-reviewer skills; for those, use Claude Code's marketplace method to install the agent-native-research-artifact category plugin.
Installation scope
Run this in your terminal. It installs these skills globally for every supported coding tool the installer detects, including Claude Code.
npx @orchestra-research/ai-research-skills install --allRun the interactive installer in your terminal.
npx @orchestra-research/ai-research-skillsSelect Install new skills, then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Claude Code and any other tools you want, then confirm Install.
Open your terminal in the project folder and run the interactive installer.
npx @orchestra-research/ai-research-skillsSelect Install to project (local), then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Claude Code and confirm Install locally. The installer copies the skills into the project's .claude/skills/ directory.
Check that the installation output lists Claude Code and its skills directory. If the new skills are not recognized, restart your agent session. Confirm that the agent can read the installed autoresearch SKILL.md before starting a research task.
Use a computer with Cursor configured, Node.js 18 or later, Git and access to GitHub. The installer detects Cursor through its existing ~/.cursor directory. Every scope below installs skill files from categories 0 and 01–21, including autoresearch. It omits the compiler, research-manager and rigor-reviewer skills; those are available through Claude Code's agent-native-research-artifact category plugin.
Installation scope
Run this in your terminal. It installs these skills globally for every supported coding tool the installer detects, including Cursor.
npx @orchestra-research/ai-research-skills install --allRun the interactive installer in your terminal.
npx @orchestra-research/ai-research-skillsSelect Install new skills, then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Cursor and any other tools you want, then confirm Install.
Open your terminal in the project folder and run the interactive installer.
npx @orchestra-research/ai-research-skillsSelect Install to project (local), then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Cursor and confirm Install locally. The installer copies the skills into the project's .cursor/skills/ directory.
Check that the installation output lists Cursor and its skills directory. If the new skills are not recognized, restart your agent session. Confirm that the agent can read the installed autoresearch SKILL.md before starting a research task.
Use a computer with Codex configured, Node.js 18 or later, Git and access to GitHub. The installer detects Codex through its existing ~/.codex directory. Every scope below installs skill files from categories 0 and 01–21, including autoresearch. It omits the compiler, research-manager and rigor-reviewer skills; those are available through Claude Code's agent-native-research-artifact category plugin.
Installation scope
Run this in your terminal. It installs these skills globally for every supported coding tool the installer detects, including Codex.
npx @orchestra-research/ai-research-skills install --allRun the interactive installer in your terminal.
npx @orchestra-research/ai-research-skillsSelect Install new skills, then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Codex and any other tools you want, then confirm Install.
Open your terminal in the project folder and run the interactive installer.
npx @orchestra-research/ai-research-skillsSelect Install to project (local), then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Codex and confirm Install locally. The installer copies the skills into the project's .codex/skills/ directory.
Check that the installation output lists Codex and its skills directory. If the new skills are not recognized, restart your agent session. Confirm that the agent can read the installed autoresearch SKILL.md before starting a research task.
Use a computer with OpenCode configured, Node.js 18 or later, Git and access to GitHub. The installer detects OpenCode through its existing ~/.config/opencode directory. Every scope below installs skill files from categories 0 and 01–21, including autoresearch. It omits the compiler, research-manager and rigor-reviewer skills; those are available through Claude Code's agent-native-research-artifact category plugin.
Installation scope
Run this in your terminal. It installs these skills globally for every supported coding tool the installer detects, including OpenCode.
npx @orchestra-research/ai-research-skills install --allRun the interactive installer in your terminal.
npx @orchestra-research/ai-research-skillsSelect Install new skills, then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select OpenCode and any other tools you want, then confirm Install.
Open your terminal in the project folder and run the interactive installer.
npx @orchestra-research/ai-research-skillsSelect Install to project (local), then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select OpenCode and confirm Install locally. The installer copies the skills into the project's .opencode/skills/ directory.
Check that the installation output lists OpenCode and its skills directory. If the new skills are not recognized, restart your agent session. Confirm that the agent can read the installed autoresearch SKILL.md before starting a research task.
Use a computer with OpenClaw configured, Node.js 18 or later, Git and access to GitHub. The installer detects OpenClaw through its existing ~/.openclaw directory. Every scope below installs skill files from categories 0 and 01–21, including autoresearch. It omits the compiler, research-manager and rigor-reviewer skills; those are available through Claude Code's agent-native-research-artifact category plugin.
Installation scope
Run this in your terminal. It installs these skills globally for every supported coding tool the installer detects, including OpenClaw. OpenClaw's skills go into ~/.openclaw/skills/.
npx @orchestra-research/ai-research-skills install --allRun the interactive installer in your terminal.
npx @orchestra-research/ai-research-skillsSelect Install new skills, then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select OpenClaw and any other tools you want, then confirm Install. OpenClaw's skills go into ~/.openclaw/skills/.
Open your terminal in your OpenClaw workspace folder and run the interactive installer.
npx @orchestra-research/ai-research-skillsSelect Install to project (local), then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select OpenClaw and confirm Install locally. The installer copies the skills into the workspace's skills/ directory.
Check that the installation output lists OpenClaw and its skills directory. Start a new OpenClaw session in the intended workspace so it discovers the installed skills. Confirm that the agent can read the installed autoresearch SKILL.md.
Use a computer with Qoder configured, Node.js 18 or later, Git and access to GitHub. The installer detects Qoder through its existing ~/.qoder directory. Every scope below installs skill files from categories 0 and 01–21, including autoresearch. It omits the compiler, research-manager and rigor-reviewer skills; those are available through Claude Code's agent-native-research-artifact category plugin.
Installation scope
Run this in your terminal. It installs these skills globally for every supported coding tool the installer detects, including Qoder. Qoder's skills go into ~/.qoder/skills/.
npx @orchestra-research/ai-research-skills install --allRun the interactive installer in your terminal.
npx @orchestra-research/ai-research-skillsSelect Install new skills, then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Qoder and any other tools you want, then confirm Install. Qoder's skills go into ~/.qoder/skills/.
Open your terminal in the project folder and run the interactive installer.
npx @orchestra-research/ai-research-skillsSelect Install to project (local), then Everything. At the target prompt, choose Select specific agents. The checkboxes start unchecked: select Qoder and confirm Install locally. The installer copies the skills into the project's .qoder/skills/ directory.
Check that the installation output lists Qoder and its skills directory. Start a new Qoder CLI session in the intended project, then open its skills list and confirm that autoresearch appears. The same skill directories are supported by Qoder IDE.
Orchestra installer source · Qoder skill directories · Qoder CLI skills