LLM-generated copy vs template-spun copy at scale

LLM-generated copy vs template-spun copy at scale

Two ways to fill the prose slots in a programmatic page. Each has a failure mode — gate accordingly.

1. Template spinning (mad-libs from data fields): cheap, deterministic, zero hallucination. Reads robotic; near-duplicate across pages. Use for: spec summaries, structured one-liners.
2. LLM generation: varied, natural, can synthesize the data into real value. Risk: hallucinated facts and silent duplication when prompts are too similar.
3. Decision rule: prose that must be factually exact and is already in your data → template it. Prose that adds interpretation/context → LLM it, then verify.
4. Guardrail for LLM: every factual claim must trace to a data field you passed in. Owner-action: ground the prompt only on supplied data, forbid outside facts.
5. Gate: run a duplication check on 100 LLM outputs — if cosine similarity is high, your prompt lacks page-specific inputs. Fix the inputs, not the temperature.

Rollback: regenerate by template id; keep prior copy versioned.

Ship gate: don't publish until all boxes are checked.