The problem
Lead Magnet AI’s conversational design was strong, but built for one business funnel: name and email capture, a waitlist hand-off. That shape only makes sense embedded in one company’s site.
The approach
Rebuilt as a standalone, six-stage skill anyone can use directly, no business behind it. Keeping what generalizes, cutting what doesn’t.
How it works
- Capture the decision
- Run the friend test
- Diagnose the block: missing information, or a permission problem (catches “I need more research” when it disguises a fear they already know the answer to)
- Check reversibility
- Deliver a personalized insight in their own words, never templated
Design decisions
- A live test caught a real bug: a three-clause question broke mid-conversation, fixed with a rule that every question stays short and single-clause
- Safety scope was checked twice: named stop conditions for self-harm and harm-to-others, red-teamed against five cases including idiom false positives
What this demonstrates
Crisis handling isn’t what the skill adds, a capable model already stops correctly on its own. What it actually changes is the shape of an ordinary conversation: one short question instead of a wall of text, at the moment someone is already overwhelmed.
100% vs. 67% pass rate across 5 evals, one run each. Full run data in benchmark.md.
Turning a business-specific product into a general-purpose tool: live iteration from real feedback, and safety design tested against both true and false positives, not assumed to work.
