Lever

The fix was deleting the step, not tuning it

Pre-generating every brief in parallel produced intermittent timeouts. First I staggered the calls, on the theory that simultaneous starts were spiking a rate limit. It helped and it did not fix it. Logging showed the real shape: the taxonomy was being serialised with fields the matching step never reads, so every call was carrying far more input than it needed against a tight per-minute token ceiling. Stripping those fields and adding prompt caching cut the payload by roughly two thirds, and under real load it still fell over. What worked was deleting the pre-generation entirely and streaming each brief on click instead: a few seconds slower to reach, and a screen that stopped timing out. Two partial fixes in a row was the signal I should have read earlier. I was making a wasteful step cheaper instead of asking whether it needed to run at all.

The tool had to start a step earlier than the sale

Most of the owners I spoke to could not yet ask build versus buy, so a tool that opens there is answering a question its buyer has not reached. Lever deliberately starts a step earlier, at what AI can do for this business at all, with complexity and a return range attached so the answer is usable rather than flattering. The bet is that meeting owners at the question they can actually ask, rather than the one a vendor wishes they were asking, is where trust in this market starts.