Code was the first big win for AI agents because code is cheap to check. The same rule decides which business processes compress next, and which stay stuck.
The mental model is wrong. You don't pour your database into a prompt. You equip the model to go look, and that's the project every business is actually working on.
Packing competing objectives into one prompt costs quality on both. The intuition for when to split is a real skill.
People who've managed teams tend to get more out of AI faster. The skills transfer directly: delegate clearly, define success, don't blame the team member for your assignment.
Modern LLMs are capable enough that step-by-step instructions constrain them. A destination outperforms a procedure. Declarative beats imperative.
You're not having a conversation with an LLM. You're writing a spec for a system that will exploit every ambiguity you leave on the table. The sooner you treat it that way, the sooner your results stop being wildly inconsistent.
AI with tools can act. What makes an agent different is what happens when the action fails.
When you pipe LLM chat responses into production systems, you're using a scratch channel as a delivery mechanism. Tool calls separate reasoning from output and fix the reliability problems you didn't know you had.
The anti-MCP take is a single-user take. When your agent serves multiple users accessing their data on remote services, you need a protocol. MCP is the one.
Bold text and CAPS in prompts aren't voodoo. They work because of how attention literally functions in transformers. Understanding the mechanism makes you a better prompt engineer.