Keystone Titles: Organize 100–500 Variants and Prevent LLM Looping (lightcapai.medium.com)

Keystone Titles reframes title generation as a clustering problem: pick one strong “keystone” title that acts as a semantic gravity well and auto-generate 100–500 related variants by identifying key intents, expanding each intent with synonyms, removing near-duplicates, scoring distinctiveness, and pruning overlapping clusters. The minimal algorithm walks through intents → synonym expansion → collision checks → keystone_variant selection → overlap pruning, producing compact clusters (e.g., theory, tutorials, tools) whose representative titles avoid cannibalization. A simple rubric (clarity, novelty, coverage, distinctiveness, memorability scored 0–2) guides keystone selection and cluster pruning, so the anchor title is both resonant for search/reader intent and distinct enough to span many useful variants. For LLM-driven ideation and drafting, the piece also tackles a practical failure mode: looping (adjacency loops, template echo, filler spirals). It prescribes six concrete loop-breakers—constraint pivot, perspective swap, anti-bigram rule, contradiction probe, timeboxing with a novelty quota, and a “what would change my mind?” pass—that serve as prompt-engineering patterns to force lexical and conceptual diversity. Together, these techniques let teams scale ideation with models while reducing redundancy, improving SEO coherence, and giving content systems deterministic controls for diversity and cluster-level distinctiveness.