Compare skills
Pick 2–4 skills and compare what really matters: fit, risk, install effort, and community signal.
Comparison matrix
Highlights show current best; tooltip explains diff/best rules.
SAS-v2.1 diff rules / risk tag notes
Start with the matrix. Open this section when you need to understand audit grades, top threats, control gaps, and best-value highlights.
Suggested baseline
Search to add skills, or paste 2–4 comma-separated slugs.
How differences are detected
A row is marked different when selected skills have distinct values. Only-differences mode hides rows that are identical.
How best values are highlighted
Audit score, evidence confidence, trust score, and community signal prefer higher values; execution risk and install friction prefer lower values.
How to read risk tags
Risk tags come from SAS-v2.1 public-evidence signals and point to command, network, secret, context, or supply-chain items to review before install.
Selected audit signals
caveman
Execution risk：High
Threat tags：unexpected code execution, identity privilege abuse, data exfiltration
Control gaps：missing license, broad permissions, shell without guardrails
Commitment Ready. 承诺。Compromiso.
Execution risk：High
Threat tags：unexpected code execution, identity privilege abuse, data exfiltration
Control gaps：missing license, broad permissions, shell without guardrails
claude-code-router
Execution risk：High
Threat tags：unexpected code execution, data exfiltration, human approval gap
Control gaps：missing license, broad permissions, shell without guardrails
| Dimension | caveman | Commitment Ready. 承诺。Compromiso. | claude-code-router |
|---|---|---|---|
| SAS-v2.1 pre-install audit | |||
Threat tags | unexpected code execution, identity privilege abuse, data exfiltration | unexpected code execution, identity privilege abuse, data exfiltration | unexpected code execution, data exfiltration, human approval gap |
Permission summary | Permission review, Network, Secrets, Command | Permission review, Network, Secrets, Command | Permission review, Network, Command |
Evidence confidence | 67% | 67% | 65% |
| Source & provenance | |||
Provenance | JuliusBrussee/caveman | openclaw/skills | musistudio/claude-code-router |
Category | Automation & Workflows | Design & Content | Automation & Workflows |
Freshness | |||
| Risk & trust | |||
Trust score | 92 | 85 | 92 |
Audit signals | |||
| Install & compatibility | |||
Supported tools | Universal | Claude, OpenClaw | Universal |
Install friction | 75 | 65 | |
| Community | |||
Stars | 27.1K | 0 | 32.2K |
Repository: LLMs-from-scratch
Author: rasbt · Source status: Clear source
Implement a ChatGPT-like LLM in PyTorch from scratch, step by step
Score basis：Clear source · Risk needs review · Universal
Repository: claude-cookbooks
Author: anthropics · Source status: Clear source
A collection of notebooks/recipes showcasing some fun and effective ways of using Claude.
Score basis：Clear source · Risk needs review · Universal
Repository: andrej-karpathy-skills
Author: forrestchang · Source status: Clear source
A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
Score basis：Clear source · Risk needs review · Universal
2026-04-13 |
2026-04-01 |
2026-03-04 |
needs credentials, network access, runs shell, writes files |
metadata-only |
Permission hints | repository clone | verify source provenance before install | repository clone |
|---|
