7/ Final point:
The bar to start has never been lower. The bar to finish is where it's always been.
Most people asking me how to start don't need a guide. They need to feel permission. Consider this it.
1/ A friend texted me asking if I knew anyone he could talk to about getting started building with AI.
The honest answer: he doesn't need anyone. Just Claude.
Here are a few questions I've gotten lately — and what I tell people.
6/ One more thing: don't quit your day job.
You can make real progress in a few hours a week while still employed.
Two caveats: check your employment agreement for IP and moonlighting clauses, and don't build anything that competes at all with your employer. Not worth it.
Honest @claudeai feedback: the upgrade that would make me a happier Claude user isn't a smarter model. It's search that actually works across my past threads.
Months of context buried in old chats, can't find it, end up re-pasting — burning tokens on stuff you already have.
Listened to @Benioff on @theallinpod describe his dream tool: something that routes AI requests to the right model based on complexity so you're not burning frontier tokens on simple tasks. He's describing @OpenRouter. It exists. Why isn't Salesforce using it?
Rolled back from Claude 4.7 to 4.6. Tool-use judgment feels off and I've caught it making obvious errors. Don't trust it the same way. Anyone else seeing the same, or has it improved since launch?
Why does every customer support system make you type your question into a chatbot, connect you to a human, and then the human asks "what's your question?" You already have it. It's right there. On the screen.