moonlight-lupin/agent-skills

Selection

Select when

• You want portable SKILL.md workflows that Claude Code can load from .claude/skills.
• You need source-grounded research, fact-checking, citation, or monitoring procedures.
• You want bundled local scripts for RAG, documents, enrichment, media, or operations.
• You prefer installing only selected skills from a broad catalog.
• You already use Hermes Agent and want its native tool architecture.
• You need EPUB or PDF ingestion into a cited semantic-search library.
• You want dry-run, confirmation, or evidence-label disciplines embedded in workflows.
• You are prepared to inspect dependencies and map Hermes-specific tools to Claude Code.

Boundaries

Avoid when

• You need the included BM25 retrieval plugin to run inside Claude Code.
• You require every skill to work without modifying Hermes-specific tool references.
• You cannot permit external API calls for skills that depend on fal.ai, PDL, Pexels, NVIDIA, or OpenRouter.
• You need a dependency-free collection across all capabilities.
• You require all components to be repository-labeled Stable.
• You need independent validation of production quality, security, or model-output accuracy.
• You require fully local processing for enrichment, media generation, or hosted embeddings.
• You want a narrowly scoped software-engineering-only skill set.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only selected first-party files were supplied, so most SKILL.md files, scripts, tests, evals, manifests, CI workflows, and dependency files were not inspected.
• Uncertainty: repository metadata is dated in 2026 relative to this analysis context, so activity, adoption, and recency cannot be independently reconciled.
• Uncertainty: repository labels such as Stable, production-tested, self-contained, and tested were not independently reproduced.
• The skill-retrieval plugin is explicitly Hermes-only and incompatible with Claude Code.
• Claude Code must map Hermes-specific tools such as web_search, web_extract, delegate_task, and cronjob where host conventions differ.
• No minimum Claude Code version, Claude compatibility matrix, or supplied Claude integration test was evidenced.
• Some tests are routing or output-contract fixtures and explicitly do not execute live models in CI.
• Skills can require third-party APIs, paid usage, credentials, network access, Python packages, curl, pandoc, browser tooling, or SQLite extension loading.
• Research, sanctions signals, media analysis, scraping, and generated outputs still require qualified human review and lawful use.
• Hosted embedding, enrichment, media, search, and stock-photo services receive submitted queries or content under their own policies.
• The library-rag files disagree with the root badge on Python requirements, stating Python 3.9+ versus repository-level Python 3.11+.
• No independent security audit, privacy assessment, comprehensive threat model, or cross-host quality benchmark was supplied.

The repository documents dry runs, confirmation gates, credential separation, local-only modes, sanctions-screening boundaries, and a nonpersistent Hermes retrieval patch. However, bundled scripts can read and write files, fetch websites, call third-party APIs, load SQLite extensions, move files, schedule work, or probe endpoints. Implementations were not fully supplied, so users should inspect each selected skill and dependency, scope credentials, review proposed filesystem changes, and preserve host approvals and sandboxing.