Tuesday morning. Vercel breach. I rotate 6 secrets in panic mode — Supabase, Stripe, OpenAI, the full set. An hour later I run npx vercel env ls just to double-check. The CLI prints 18 more secrets in cleartext right in my terminal. In audit mode. Read-only, the docs said. And I had trusted that. (Full post-mortem)
This week
StackOverflow dropped the number of the year: 84% of developers use AI daily. 29% actually trust it. The gap between those two is exactly what I just lived through with Vercel — we lean on tools we've stopped verifying, because they've slid into reflex territory.
Everyone talks about the review skill — read the diff, correct the output, validate before merge. Not wrong, but that's step 2. The step 1 that's disappearing is refusal. Saying "no, redo" when the output looks fine. Saying "no, I'll write this part myself" when the flow pulls you in. Refusal is the only place judgment builds. Acceptance grows nothing.
Here's the test that works: tomorrow morning, count how many AI outputs you accept while scrolling (a diff, an improved prompt, an answer) versus how many you refuse. If the ratio dips below 1:5 (one refusal for every five accepts), the muscle is atrophied. It rebuilds in 2–3 weeks, no longer — but only if you count.
In the news
Vercel breach (April 21) — integration tokens leaked via an internal endpoint. Official postmortem: "no customer secrets exposed." Technically correct. In practice, npx vercel env ls still prints your "Plain" env vars in cleartext — if you didn't tick Sensitive at creation, they're still decryptable on Vercel's side. Mandatory rotation this weekend. The breakdown + checklist.
This week on the blog
Gumroad Just Became Scriptable. The 12-Agent Stack That Runs My Store. — Gumroad's CLI/API shipped. I replaced 6 SaaS tools with 12 local agents.
Every AI Got Dumber in 30 Days. And It's About to Get Worse. — Compute allocation shift. Why GPT-5 / Claude / Gemini have been regressing since March.
Opus 4.7 Didn't Break Your Prompts. You Just Missed the 2 Rewrites That Matter. — Short migration guide, 2 patterns to rewrite, zero bullshit.
Phil
PS: when's the last time you told Claude/GPT "no, redo" on an output that looked fine? Hit reply with the concrete example — I'm collecting them for a follow-up.
PPS: what's your biggest current problem with AI? One-line reply — I read everything.
You're receiving this because you signed up at rentierdigital.beehiiv.com.