Suggest edit — Reviewers & Registries (meta)

Title

Name

Note

---
title: Reviewers & Registries (meta)
visibility: public
tags: [meta, claude, skills, trust-network, reviewers]
---

# Meta: Trusted Claude Skills + Skill Reviewers

This is the **meta page**: not the skills themselves, but **who reviews and vouches for them**, plus the skill registries / discovery surfaces that exist as of April 2026. Use this as the starting node in a "who do I trust" graph for Claude Code skills.

> See also [[index|Trusted Claude Skills]] for the actual catalog, and [[@jacobcole/curation-trust-network/index|Curation & Trust Networks]] for the broader pattern.

---

## Tier 1 — Authoritative

| Source | What | Why trust |
|---|---|---|
| [Anthropic Engineering blog](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills) | Original Skills design + threat model | Defines the spec, knows what it's trying to constrain |
| [`anthropics/skills`](https://github.com/anthropics/skills) | Official skill repo | First-party. Default-trust. |
| [agentskills.io](https://agentskills.io/home) | Open standard, spec docs | Maintained as the public spec entry point |

## Tier 2 — Security-vetting registries

| Source | Vetting model | Caveats |
|---|---|---|
| [Skills Directory](https://www.skillsdirectory.com/) | Every skill scanned for malware, prompt injection, credential theft | Static scanning misses behavioral risk |
| [Tech Leads Club agent-skills](https://github.com/tech-leads-club/agent-skills) | "Verified, tested, safe"; published the 13% malicious finding | Smaller catalog; opinionated curation |
| [SkillHub](https://www.skillhub.club/) | AI-evaluated on 5 dims; S-rank (9.0+) only | LLM-evaluated quality ≠ security audit. S-rank is signal, not proof. |
| [LobeHub Skills](https://lobehub.com/skills) | "Security-first" red-flag/permission/pattern checks | Self-reported standards |

## Tier 3 — Discovery only (no vetting)

| Source | Notes |
|---|---|
| [SkillsMP](https://skillsmp.com/) | Aggregates GitHub skills. Discovery surface; **assume untrusted**. |
| [claudeskills.info](https://claudeskills.info/skills/) | "140+ free open-source skills." No audit gate. |
| Random GitHub repos via search | Unaudited by definition |

---

## Reviewers Jacob trusts personally

These are individuals / orgs whose endorsement of a skill carries weight:

- **Anthropic Applied AI team** — anything they ship in `anthropics/skills`
- **Tech Leads Club** — they actually published the malicious-skills audit data; signal that they've done the work
- _(Add as discovered — open node)_

## Reviewers we are watching but not yet trusting blindly

- LobeHub — security-first claim is unverified; treat their "verified" badge as a soft signal
- SkillHub S-rank curators — quality dimension is well-defined but security is a separate axis

## Empty seats — open invitations

- An independent academic security researcher publishing a regular Skill audit would be highly trusted. **Doesn't appear to exist yet.**
- A Mozilla-style open-foundation review board for skills. **Doesn't exist yet.**
- A community "reproducible audit" registry where multiple reviewers must concur on a skill's safety. **Open opportunity.**

---

## Trust hierarchy (decision flow)

```
Need a skill?
  └─ Available in anthropics/skills? ───────────────── ✅ use it
  └─ Available + you wrote it? ────────────────────── ✅ use it
  └─ Available + Tech Leads Club verified? ────────── ✅ use it (after personal read)
  └─ Available in Skills Directory + SkillHub S-rank? ─ 🟡 use after sandbox test
  └─ Random marketplace ──────────────────────────── 🔴 audit code, sandbox, then maybe
  └─ Random GitHub link ──────────────────────────── 🔴🔴 default-deny
```

## What this meta page should become

- A dynamic registry where each skill in [[index|Trusted Claude Skills]] links to **which reviewer(s) blessed it** and **on which date**
- A "trust chain" record per skill (who → who → who) so trust is transitive but auditable
- Comments / disputes from other WikiHub users about specific reviewers

That's the curation use case. See [[@jacobcole/curation-trust-network/index]].