This week a fascinating report came out.
But the most important finding in @edwin's research is the one nobody is talking about.
And it's not the tool rankings.
Build vs. Buy
The report tracked 2,073 tool picks across 20 categories.
Custom/DIY appeared in 12 categories with 252 total picks. More than any single product.
Claude's default instinct isn't “use PostHog or LaunchDarkly.”
It's “I'll just build Feature Flags myself.”
Winners today. Losers tomorrow
The other finding nobody is connecting:
Prisma went from 79% to 0% between model versions
Redis from 93% to 29%
These look less like preference shifts and more like extinction events.
Not because the product got worse, but because the training data shifted.
LLMs have Recency Bias
Newer models seem to favor newer tools.
The implication is wild: your adoption curve now has to outrun the model retraining cycle.
If you're not generating fresh content, you're fading from the AI's memory.
But if you keep shipping, the same cycle works in your favor.
The Growth is Real
At @resend we're seeing this firsthand.
Claude Code recommends us 62.7% of the time for email.
Our internal data shows that, of the people who say where they found Resend, 1 in 4 say Claude.
This doesn't translate to just new sign ups. It translates to actual revenue.
But we're not taking this for granted.