Orchestra-Research/AI-Research-SKILLs

Selection

Select when

• You want Claude Code guidance spanning both research reasoning and ML engineering.
• You need several framework-specific skills from one installable collection.
• You want category-level or individual skill selection.
• You want an autoresearch workflow that routes across specialized skills.
• You need guidance for fine-tuning, RL post-training, distributed training, evaluation, or serving.
• You need mechanistic-interpretability workflows alongside training guidance.
• You want paper-writing, LaTeX-template, and academic-plotting support.
• You want project-local skills that can be version controlled.
• You use multiple SKILL.md-compatible coding agents.
• You want research outputs organized into auditable agent-native artifacts.

Boundaries

Avoid when

• You need independently verified scientific conclusions rather than agent guidance.
• You need a turnkey compute environment with dependencies and credentials already configured.
• You want a small domain-neutral Claude Code extension.
• You cannot review generated code, experimental design, citations, or claims.
• You require offline-only operation for skills that reference remote APIs or services.
• You need guaranteed compatibility with a specific framework or host version.
• You require an independently audited installer or skill corpus.
• You do not want an autonomous agent writing files or running experiments.
• You need deterministic reproduction of the showcased autonomous demos.
• You require consistent inventory details with no documentation drift.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only selected first-party files were supplied, so the 98 SKILL.md files, installer source, marketplace manifest, CI workflows, optional scripts, and most references were not directly inspected.
• Uncertainty: metadata and release notes are dated in 2026 relative to this analysis context, so activity, version, adoption, and recency claims cannot be independently reconciled.
• The root README contains stale 87-skill statistics despite stating that inventory was reconciled to 98 skills, indicating residual documentation drift.
• The package category examples disagree with the root inventory on model-architecture and fine-tuning counts and named inclusions.
• Repository claims of production-ready, battle-tested, autonomous, comprehensive, or research-grade quality were not independently validated.
• Demo conclusions and performance numbers are repository-reported, sometimes use synthetic data, and do not establish general scientific reliability.
• Autonomous literature review, experimentation, causal interpretation, citation handling, and paper writing still require expert human oversight.
• Many skills depend on third-party frameworks, GPUs, cloud services, API keys, datasets, system packages, or paid model access.
• Optional helper scripts can execute code or transmit inputs to external services, and their implementations were not supplied for inspection.
• No minimum Claude Code version, host compatibility test matrix, telemetry disclosure, threat model, or independent security audit was supplied.
• The installer downloads skills and creates symlinks or copies in multiple agent directories, which changes local configuration and expands the review surface.
• Individual referenced libraries and bundled conference materials may have licenses, versions, deadlines, or requirements distinct from the repository's MIT license.

Evidence: release notes state that installer dependencies were pinned to patched versions and that npm audit reported zero vulnerabilities; this was not verified from manifests or CI output. The installer downloads content and writes symlinks or copies into agent directories, while optional skill scripts may execute code or use external APIs. Users should inspect installed skills, scripts, templates, dependency versions, and data flows before use; no independent audit or threat model was supplied.