Agent Skills
› mtarcure/claude-vibe-squad
› keyword-clustering
keyword-clustering
GitHub将搜索查询按意图和主题聚类，映射至页面并生成包含来源、日期、本地化等元数据的证据表。确保规则可复现，明确分裂/合并逻辑及例外原因，避免内容重复竞争。
触发场景
需要对大量搜索查询进行意图分类和分组
需要为SEO优化生成关键词到页面的映射关系
需要分析现有或潜在的内容重叠（Cannibalization）
安装
npx skills add mtarcure/claude-vibe-squad --skill keyword-clustering -g -y
SKILL.md
Frontmatter
{
"name": "keyword-clustering",
"audience": "specialist",
"description": "Use when supplied or grounded search queries must be grouped into reproducible intent\/topic clusters and mapped one-cluster-per-page; emit the evidence table with source\/date\/locale\/device and explicit split\/merge rules. Not for measuring ranking impact without analytics."
}
Keyword Clustering
Group supplied or grounded search queries by intent and topic, and map each reproducible cluster to a page.
Required evidence table
Produce one row per normalized query with these fields: source, collection date, locale, device, raw query, normalized query, intent label, intent confidence, similarity/split-merge rule applied, cluster ID, target page, and exception rationale. Pin the similarity method and split/merge threshold before clustering; if judgment overrides that rule, the row's exception rationale makes the override reviewable.
Steps
- Record the query source, collection date, locale, and device; never invent query or volume evidence.
- Normalize each raw query with a stated rule while preserving the raw value in the table.
- Tag intent as informational / navigational / transactional / commercial, with a confidence value and an explicit multi-intent exception where one label would misrepresent the query.
- Apply the pinned similarity and split/merge rule so one cluster expresses one reviewable user need.
- Map each cluster to one target page/content piece and record why any mapping exception is necessary.
- Flag cannibalization wherever existing or proposed pages compete for the same cluster.
Acceptance
- Every query has complete source/date/locale/device provenance and an intent confidence.
- The normalization, similarity, and split/merge rules are fixed and replayable; exceptions carry reasons.
- Each cluster maps to exactly one page (no two pages target the same cluster).
- Queries/volumes are grounded, not fabricated, and the required evidence table accompanies the map.
版本历史
- d5262e2 当前 2026-09-11 11:24
