You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
S3 generates scripts in a vacuum — no niche, no audience, no topic constraints. Output is generic ("phone battery hack", "bad morning story") because the prompt has no context about what the creator's channel is about or who watches it.
This is the biggest quality gap in the pipeline.
Suggested Improvements
Add niche and target_audience to the generation prompt (from creator profile)
Pass 2-3 verbatim hook examples from S2 as few-shot references
Add topic selection as an explicit step — from creator input or inferred from past content
Add "specificity score" self-evaluation — force the LLM to rate how generic its output is
Generate hooks independently first, then build scripts around the best hooks (hook quality is the Add spec v2 design document #1 driver of views)
Context
See design/pipeline-quality-review.md — S3 is listed as priority #1 for quality improvements.
Problem
S3 generates scripts in a vacuum — no niche, no audience, no topic constraints. Output is generic ("phone battery hack", "bad morning story") because the prompt has no context about what the creator's channel is about or who watches it.
This is the biggest quality gap in the pipeline.
Suggested Improvements
nicheandtarget_audienceto the generation prompt (from creator profile)Context
See
design/pipeline-quality-review.md— S3 is listed as priority #1 for quality improvements.Depends on