Last Wednesday night, I froze for a few seconds when this hit my feed
Claude designed working protein molecules on its own — not simulated, actually synthesized in a lab. 13 out of 14 targets worked.
I've spent years in consulting selling "expert judgment." If this becomes normal, I need to rethink what I'm even selling.
What actually happened, and who's already using it
Anthropic CEO Dario Amodei said just last week that "a few more months until sparks fly." Four days later, sparks flew. They gave Claude (Mythos Preview and Opus 4.8) an expert-written prompt, plus web access and tool permissions, and let it run protein design solo. The actual lab work went to Twist Bioscience and Adaptyv Bio — AI only handled the design.
Reading this reminded me of a nutrition brand client I'll call Lao Zhou. He paid 8,000 RMB for a formulator who basically just iterated by experience. I had wondered back then whether AI could brute-force the combinations a formulator tries. Nobody built it, that's all. Kicking myself for not digging deeper.
Even wilder: Opus 5 reads raw instrument files on its own, hits 96.4% purity in 19 minutes — the lab's own report takes 4 days.
Can you replicate this today? Costs are low
We can't replicate protein design (no need to), but the pattern of "AI doing the expert's work" — that's ours to use:
Money: Claude paid tier is $20/month (around 145 RMB). Free tier handles most of it.
Time: Sign up plus first real prompt — 30 minutes.
Tech barrier: No coding, just typing. The only thing that matters — write prompts like job postings. Background, goals, constraints — five sentences minimum.
First step: Open claude.ai, paste in a "needs expert judgment" task you've been stuck on, with enough background.
Three types of people, three different next moves
Just starting out: Stick with free tier. Pick the thing that hurts most — writing quotes, doing client research, drafting contracts — let Claude ship v1. Editing beats writing from scratch by 3x.
Got 1–2 stable clients: Spend an afternoon building an "expert prompt library" — one high-quality prompt per client per frequent question. This is your moat. The earlier you build it, the more it's worth.
Scaling up: Pull up the expense sheet. Look at what flows monthly to lawyers, designers, consultants. Which of those can AI handle 80% of, and I review the last 20%? I tried this on contract review — AI flagged three traps I almost walked into.
One last thing: this tool isn't for everyone. If AI can't help your current work yet, no need to panic. Serve clients well first.