Engineering Blueprint
Safety checkedConduct AI Research End to End With Skills
98 AI research engineering skills spanning the full research lifecycle from idea to paper and beyond across 23 functional domains.
Install Command
Run this command to deploy the blueprint to your environment.
What problem does this solve?
Conducts autonomous AI research from literature survey through experiment execution to paper writing by orchestrating 98 specialized skills across model architecture, training, evaluation, interpretability, and publication. Eliminates the burden of mastering dozens of fragmented tools by routing research tasks to domain-specific skills that handle implementation details, troubleshooting, and production workflows.
How does it work?
Once installed, load the autoresearch orchestrator, which manages the full research lifecycle using a two-loop architecture. The orchestrator accepts a research question, runs literature surveys and ideation through domain skills (model training, evaluation, mechanistic interpretability, etc.), executes experiments using framework-specific skills, and synthesizes results into papers and research presentations. Output: end-to-end research with context carried between stages.
What's the biggest win?
Agents spend time testing hypotheses and iterating on ideas instead of debugging infrastructure, accelerating the pace of scientific discovery.
What should I know technically?
Skills are installed via npm package (@orchestra-research/ai-research-skills) to ~/.orchestra/skills/ with auto-detection of your coding agent (Claude Code, OpenCode, Cursor, Codex, Hermes, Qwen Code, Gemini CLI). Installation falls back to file copy on Windows if symlinks unavailable. Requires Node.js for npm-based installation; individual skills support Python 3.8+, PyTorch 2.0+, and CUDA 11.8+ depending on the framework.
What should I watch out for?
Requires understanding which skill to invoke for your specific task (mitigated by autoresearch orchestrator routing). Some domains like distributed training require infrastructure knowledge. Individual skills may have incompatibilities with certain framework versions.
Key Features
98 Comprehensive Skills
Production-ready guidance covering architectures, training, evaluation, safety, optimization, inference, agents, RAG, multimodal, and emerging techniques.
Two-Loop Orchestration
Autoresearch skill manages full lifecycle: inner loop optimizes experiments, outer loop synthesizes and writes papers automatically.
23 Functional Domains
Parallel capability areas covering model architecture, fine-tuning, distributed training, inference serving, mechanistic interpretability, and more.
Research Artifact Generation
Outputs research presentations (HTML/PDF), findings tracking, session provenance, and falsifiable agent-native artifacts.
Battle-Tested Production Workflows
Each skill sourced from official docs, real GitHub issues, and production-proven techniques with 300-500 lines of expert guidance.
Tools in this Blueprint
About This Blueprint
- License
- MIT
- Industry
- Technology
- Skills
- 1 workflows, 0 sub-skills, 98 standalone