playbook-creative-ideas

Define three fixed idea slots up front in the seller's own words, then have the model fill each slot only with a detail true of that company. Slot-defined generation scored 21/23 usable versus 2/5 for free-form ideation.

Require the model to output an evidence quote alongside each bullet, then assert the quote appears as a real substring of the input evidence after normalization. A bullet whose evidence is not in the input was invented and gets blanked, with no second model call needed.

Use gpt-4o-mini inside Clay at roughly $0.27 per 1,000 rows. A nano-class reasoning model with reasoning unset burns hidden reasoning tokens, costs about 19x more, and frequently returns blank content; outside Clay, a nano model at minimal reasoning costs $0.12 per 1,000 rows.

Gating each source on the previous evidence being empty stops at the first non-empty result, which is often a two-sentence boilerplate. Gate on length instead (under 200 characters) so the waterfall continues until it finds substantive evidence or abstains.

Any empty bullet excludes the entire row and routes it to a separate non-ideas campaign. A 3-bullet email with 2 bullets is a different email, and sequencers cannot pin a lead to a specific sequence variant.

Skip it when one personalized sentence atop a fixed email suffices, when the hook is an event, or when the operator cannot name three fixed slots. An idea the operator did not ask for is one the seller cannot deliver on the call.