lythoskill-curator

Files (18)

• SKILL.md
• architecture.md · references/
• catalog-db.md · references/
• combination-patterns.md · references/
• design-principles.md · references/
• frontmatter-fields.md · references/
• mcp-discovery.md · references/
• recommendation-workflow.md · references/
• registry-schema.md · references/
• why-this-shape.md · references/
• zk-subagent-patterns.md · references/
• judge.md · scripts/bdd/curator-find-fuzzy-bdd/
• reproduce.sh · scripts/bdd/curator-find-fuzzy-bdd/
• judge.md · scripts/bdd/curator-find-repo-bdd/
• reproduce.sh · scripts/bdd/curator-find-repo-bdd/
• judge.md · scripts/bdd/find-hit-miss/
• reproduce.sh · scripts/bdd/find-hit-miss/
• curator.sh · scripts/

SKILL.md (first lines)

# Skill Curator
> Curator's perspective (买家秀 "buyer show" — what hands-on use actually found,
> versus the author's 卖家秀 "seller show" claims) = card searcher + personal notes
> + composition taste.
> CLI is mechanical glue (scan, query, tag, audit). Agent does the thinking.

> **Invocation.** `curator <cmd>` below is shorthand for
> `bunx @lythos/skill-curator@{{PACKAGE_VERSION}} <cmd>` (no install), or
> `lythoskill-curator <cmd>` if installed globally. The bare name is NOT a
> registered bin — in-repo alias convention only. `deck <cmd>` / `arena <cmd>`
> map to `@lythos/skill-deck` / `@lythos/skill-arena` the same way.
> Want the shorthand? `alias curator='bunx @lythos/skill-curator@latest'` — optional.

## Mental Model: Curator = the Curator's Perspective (策展者)

Curator is NOT a discovery engine. It's the **curator's personal knowledge base** —
a card searcher + personal notes system for the skill ecosystem.

**Two complementary modes**:

| Mode | Metaphor | What it does |
|------|----------|-------------|
| Journalist (记者) | Investigation + narrative + expression | Fact-check claims, cross-reference sources, detect bias, assign confidence |
| Architect (架构师) | Composition aesthetics | Understand synergies, deduce combos, judge structural fit, design archetypes |

**Composition taste (组卡审美) has three inputs**:
1. **Arena field data** — quantitative: scores, pass/fail, performance data
2. **Design critique** — qualitative: your own judgment of a skill's design, clarity, fit
3. **Combo reasoning** — systemic: how skills compose, what synergies emerge, what archetype they form

A curator doesn't just verify facts (Journalist mode). A curator understands **how skills
combine to be beautiful** (Architect mode). This is what separates a card database from
a deck builder.

**Curator builds on deck + arena**: it never evaluates a skill in isolation — it evaluates
"this card, in this deck, on this task." Arena tests deck-level capability; curator records
deck-task-level QA. A card that excels in one deck may be useless in another; the Architect
mode exists to reason about exactly this context dependence.

## Discovery SOP (Agent-Driven)

The explore slot is dominated by **agent + search** (gh CLI first, WebSearch as fallback).
Curator's job is NOT to be the discovery engine — it's the **local cache** that makes
discovery faster, and the **enrichment layer** that remembers what was found.

```
1. curator find <bare-name>                                   ← local cold pool: "already have it?"
2. curator query "SELECT name, description FROM skills        ← local cache: "anything similar?"
   WHERE description LIKE '%<keyword>%'"
3. gh search code "<bare-name>" --filename "SKILL.md"         ← GitHub code search: precise, hits exact name
4. gh search repos "<keyword>" --topic "agent-skills"         ← GitHub topic search: discover skill repos
5. WebSearch for "<name> skill site:github.com"               ← web fallback (site: prefix when supported)
6. WebFetch / gh CLI to inspect candidates                   ← deep dive
7. curator add <locator> --pool ...                           ← seed cold pool
8. curator scan                                               ← re-index
9. curator tag <name> --niche "<classification>"              ← agent-enriched metadata (L3)
   [--qa '{"source_type":"self/arena","signal_value":8,...}']
10. arena single/vs                                            ← test before adopting
11. curator tag <name> --qa '{"source_type":"self/arena"...}'  ← record test results
12. Recommend with confidence: "skill X fits because...        ← agent reasoning
    (3 arena PASS + hub A confirms + curator scan clean)"
```

**Curator is NOT the discovery engine.** It's the agent's local data source.
The agent combines curator query + gh search + WebSearch + its own reasoning for
discover → rank → recommend.

### Three Discovery Scenarios

| Scenario | What you know | Start here |
|----------|--------------|------------|
| "I know the skill name" | bare name (e.g., "fullstack-dev") | `curator find` → gh search code |
| "I know the person/org" | fuzzy name (e.g., a Chinese creator's online handle) | WebSearch → gh search code for precise path |
| "I know the repo URL" | repo URL (e.g., `github.com/lijigang/ljg-skills`) | gh api peek → curator add |

### Search Precision Ladder

When `curator find` misses, use these in order — fastest to broadest:

| Priority | Method | Speed | Precision | Example |
|----------|--------|-------|-----------|---------|
| 1 | `gh search code "<name>" --filename "SKILL.md"` | ~2s | Highest | `gh search code "fullstack-dev" --filename "SKILL.md"` |
| 2 | `gh search code "<name> skill" --filename "SKILL.md"` | ~3s | High | broader match |
| 3 | `gh search repos "<topic>" --topic "agent-skills"` | ~3s | Medium | `gh search repos "writer" --topic "agent-skills"` |
| 4 | WebSearch `"<name> skill site:github.com"` | ~5s | Medium | site: prefix for domain filtering |
| 5 | WebSearch `"skill <name>"` | ~5s | Broad | without site: prefix |

`gh search code` is the most effective because it searches file CONTENTS — a skill named
"fullstack-dev" always has `name: fullstack-dev` in its SKILL.md frontmatter. This is
more precise than repo search or web search.

### Repo Exploration — "I know the repo, what's inside?"

Common social-media discovery pattern: someone shares a GitHub repo URL, but you don't
know what skills it contains. Common for monorepos (e.g., `JimLiu/baoyu-skills` has 22,
`lijigang/ljg-skills` has 21). Use `gh api` to peek without cloning:

```bash
# 1. List top-level files/dirs in the repo
gh api repos/<owner>/<repo>/contents --jq '.[].name'

# 2. If there's a skills/ dir, list it
gh api repos/<owner>/<repo>/contents/skills --jq '.[].name'

# 3. Optionally peek at specific SKILL.md frontmatter
gh api repos/<owner>/<repo>/contents/skills/<name>/SKILL.md --jq '.content' | base64 -d | head -10
```

**Real example** — discovering what's in `lijigang/ljg-skills`:
```
$ gh api repos/lijigang/ljg-skills/contents/skills --jq '.[].name'
ljg-book  ljg-card  ljg-invest  ljg-learn  ljg-paper  ljg-plain
ljg-present  ljg-push  ljg-qa  ljg-rank  ljg-read  ljg-think
ljg-travel  ljg-word  ljg-writes  ...  (21 skills total)

$ gh api repos/lijigang/ljg-skills/contents/skills/ljg-think/SKILL.md \
    --jq '.content' | base64 -d | head -5
---
name: ljg-think

From lythos-labs/lythoskill, MIT.

More from lythos-labs/lythoskill