Suggest edit — Use Cases (deep)

Title

Name

Note

---
title: Use Cases (deep)
visibility: public
tags: [curation, trust-network, use-cases]
---

# Curation × Trust Network — Use Cases (deep)

A long-form companion to [[index|the meta wiki]]. Each use case below is rated against the 6 structural criteria (abundance / risk / heterogeneity / no-single-arbiter / transitivity / audit-trails). Scale: ★ to ★★★★★.

---

## Software & agent ecosystems

### Claude Code Skills — ★★★★★
- **Abundance:** 7,000+ skills (SkillHub) and growing weekly
- **Risk:** Code-execution on dev box; data exfil; credential theft
- **Quality variance:** 13% with critical vulns (Tech Leads Club); also some genuinely excellent
- **No single arbiter:** Anthropic explicitly cannot vet every community skill
- **Transitivity:** Recommendation chains work well ("X uses Y because Z vouched")
- **Audit:** Static scan helps (Skills Directory) but behavioral audits are rare
- **Catalog:** [[@jacobcole/trusted-claude-skills/index]]

### OpenClaw Skills — ★★★★★
- 5,400+ skills, **820+ confirmed malicious** in audit
- Cryptominers, credential harvesters, exfil patterns
- Default-untrusted is the only safe stance
- **Catalog:** [[@jacobcole/trusted-openclaw-skills/index]]

### MCP Servers — ★★★★
- Growing fast; many implementations of the same protocol
- Each MCP server can issue tool calls; full process trust
- No verified registry yet; provenance often murky
- **Open opportunity** for a curated MCP-server trust catalog

### AI coding agents (Claude Code, Cursor, Aider, Codex CLI, Antigravity, OpenCode, Continue.dev, etc.) — ★★★
- Many overlapping ones, switching cost is days
- Risk is mostly time/lock-in, not security (these are themselves trusted at install)
- Trust-net signal: which devs you respect actually use which one daily

### Browser extensions — ★★★★★
- The original case. Same-origin power, vast catalog, occasional malware
- Chrome Web Store does some curation; not enough
- People still reuse trust-net cues: "this one is by [respected person]"

### VS Code / Cursor extensions — ★★★★
- Run inside the IDE, can read all open code
- Marketplace has done some vetting but high-trust extensions are still community-known

### Open-source LLM model weights — ★★★
- Hundreds of variants; possible backdoors; license-status varies
- Hugging Face is the registry; trust signals are model-card author + benchmarks
- Emerging audit space (`safetensors` provenance, weight-fingerprinting)

---

## Health & body

### Bodywork practitioners (chiro, PT, ART, Rolfing, qigong) — ★★★★★
- Local, heterogeneous, hard to evaluate from outside
- Wrong choice = months of pain or worse
- **Trust-net works extremely well** ("my friend Alice cured her RSI with X") — most existing solutions are wrong-modality (Yelp reviews don't capture this)
- See [[@jacobcole/health-log|Health Log]] (private)

### Supplements / longevity protocols — ★★★★
- Vast literature, lots of hype
- Most reviewers are conflicted (selling products) or unqualified
- Need a trust net that flags conflict-of-interest disclosures explicitly

### Mental-health / contemplative teachers — ★★★★★
- Years of life invested per choice
- Cult dynamics are a real failure mode
- Reputation networks already exist informally; making them legible helps a lot

### Diet / nutrition advice — ★★★★
- Same shape as supplements but worse — fads, ideology, expert disagreement

---

## People

### Investors / VCs — ★★★★★
- The existing trust net (warm intros, founder backchannels) is exactly this pattern, just opaque
- A semi-private trust-graph wiki is huge value
- See [[@jacobcole/investors-wiki|Investors Wiki]] (private)

### Founders / collaborators / hires — ★★★★★
- "Should I work with this person?" — high stakes, no central truth
- Reference-checking is informal trust-graph traversal
- Risk: reputational liability of writing things down

### Service professionals (lawyers, accountants, doctors) — ★★★★
- Yelp doesn't work; word-of-mouth does
- Trust net + segment-by-need is the right shape

---

## Information & culture

### Books to read — ★★★
- Stakes are time, not money/health
- Trust-net (Goodreads-friends, blog recommendations) already works decently
- Curation value is sorting **for a specific need state**, not "best ever"

### News / commentary sources — ★★★★★
- Bias and ownership matters; readers usually can't tell
- Existing efforts: media-bias-fact-check, allsides
- Strong trust-net case: who in your circle reads which sources, and why
- See `lists-to-curate` items #16 (media bias map) and #14 ("good commentaries")

### Academic papers / fields — ★★★★
- Trust net = whose reading list do I borrow?
- Citation graphs are an implicit trust net but lossy

### Quotes / aphorisms / wisdom — ★★
- Lower stakes; mostly aesthetic
- See [[@jacobcole/quotes|Favorite Quotes]]

---

## Local / activities

### Bay Area activities — ★★
- Lower stakes, low malice risk
- Trust net useful but soft

### Restaurants / venues — ★★
- Lower stakes, but algorithmic curation has degraded badly
- Friend-recs > Yelp consistently

### Events / community houses — ★★★
- Higher stakes (you're committing time + money + presence)
- Friends' attendance signals are gold

---

## Opportunities & alerts

### Contests / awards / grants — ★★★
- Many exist, hard to find the right one
- See `lists-to-curate` item #24 (Berggruen + others)

### Fitness benchmarks — ★★★
- Many proposed; few rigorous
- See `lists-to-curate` item #7

---

## Patterns observable across all these

1. **Algorithmic recommendation engines lose to trust nets** when (a) stakes are real and (b) the long tail dominates.
2. **Existing trust nets are mostly informal and lossy** — phone calls, DM'd intros, Slack "any recs?" threads. Making these legible is the win.
3. **Privacy varies by case** — investors and people-graphs need privacy; books and skills don't.
4. **Negative endorsements (denylists) are rarer and more valuable** than positive ones. AI Makers' "5 OpenClaw skills to avoid" is exemplary.
5. **The reviewer is the unit of trust, not the review.** A good curation system makes reviewers identifiable and stake-holding.

---

## What WikiHub should support to serve these cases

- **Per-page endorsements** with reviewer identity and timestamp (signed if possible)
- **Reviewer profile pages** with their endorsement history, expertise tags, conflicts disclosed
- **Reusable trust assertions** (page → page) so traversal is a graph query, not a re-read
- **Visibility scoping** so private trust nets (e.g. investor wiki) work alongside public ones
- **Revocation as a diff** (git already gives this for free)
- **Denylists** as first-class citizens, not afterthoughts