Nothing goes out that can't be defended.
Every pitch hadef writes passes five checks before it reaches your review queue. Quote accuracy, URL liveness, specificity, quote ownership, and confidence scoring. Most AI writing tools hope you don't look closely. hadef is built for hosts who do.
The five checks
Quote accuracy
Does the hook quote appear verbatim in the source transcript? Substring match, not semantic. Miss by one word and it gets flagged for manual review.
URL liveness
Does the source link return HTTP 200? Dead links don't ship. Paywalled links are marked, not hidden.
Specificity
Could this pitch be sent to a different guest with only minor edits? If yes, QA fails and the pitch goes back for a rewrite.
Quote ownership
Is the hook quote spoken BY the guest, not ABOUT the guest? Third-party mentions never become hooks.
Confidence score
Transcript coverage strong? Quote verified direct? If either is weak, the pitch ships with a MANUAL_REVIEW flag instead of green-lit.
Three conditions
Zero valid transcripts plus a hook quote present. Quote marked MANUAL_REQUIRED. Concept confidence LOW. Any one blocks the auto-send path.
What the QA record looks like
Every pitch ships with a QA object. Five booleans, one failure reason, one recommendation. Stored in Supabase, surfaced in the pipeline view, exported on request. When you mark a pitch sent, the QA record travels with it.
Reply rates correlate with QA pass rates. Pitches that pass all five checks reply at five to seven percent. Pitches with any soft flag drop to two to three percent. Don't ship what can't clear the bar.
See QA in action
Run a free pitch. See every check, every result, every reason the engine either greenlit it or flagged it.