Ruben

Renamed myself Claude in Slack this morning. Three people have already thanked me for my thorough analysis. One asked me to stop being so agreeable.

Every UUID in your prompt costs ~20 tokens and carries zero information the model can use. It's the most expensive way to say "thing number 7". Short slugs beat random hex. Design your IDs for the reader, and these days the reader is a tokenizer.

Claude Sonnet 5 pricing is “unchanged” vs 4.6 on paper… …but the new tokenizer produces ~30% more tokens for the same text. That means both input and output costs can effectively rise ~30% once intro pricing ends. Benchmark cost per successful task, not price per token.

Someone built a site that scans your SaaS and tells you how likely it is to be replaced by a Claude Skill. Which is, quite literally, a markdown file. If your product is a workflow layer on top of something an LLM can do natively... what's your moat? deathbyclawd.com

Vibe coding introduced a problem I didn't have before: constant micro context switches. Every prompt has a 30-90 second gap. Too short to do real work, too long to sit still. So you check something, open another tab, lose your train of thought.

The vibe gap: productivity up, understanding down, vibes immaculate.

TIL batch processing for LLMs is basically "I don't need this now, just get back to me" and it costs half as much. Why was I running everything real-time again?

If you're building with LLMs and not using prompt caching, you're literally paying to re-read the same book every single API call. Cache your system prompts & context. Cut costs up to 90%. It's one of the most underrated optimizations out there.

PSA: Claude Cowork quietly eats ~12 GB of disk space when enabled. That's a 10.8 GB VM bundle just sitting in ~/Library/Application Support/Claude. Might want to keep an eye on that if you're running low on storage. 👀

Looking for a domain for your AI startup? These emoji-inspired .ai domains are still available: 🗝️ oldkey.ai 🎲 gamedie.ai 🌊 waterwave.ai 😠 angryface.ai ⛽ fuelpump.ai Check out: emoji-domains.com

Indie hackers celebrating $10k MRR on a 3-week-old AI wrapper is like celebrating a marriage on the first date. At least wait for the churn data before you tattoo that number on your Twitter bio.

The "1% better every day = 37x in a year" crowd ignores that going from 90 to 91 is ten times harder than going from 10 to 11. Growth isn't compound interest. It's plateaus you break through by changing HOW you improve. The formula is a lie. The habit isn't.

v0 pro tip: iterate on UI with dummy data first. Real DBs and CRUD add complexity—and credits burn faster.