Recap

We're three years into a five-hundred-year skill. About 1200 ELO. When the model lets you down, the closed-minded blame the AI; the open-minded ask what they could have done differently. The second is the faster path, by a lot.

When the output disappoints, walk the four levers. Better tool. Better model. Better prompt. Better context. The biggest lever is almost always context.

Practice a repertoire of techniques, like:

• What questions should you ask me before starting?
• What's the smallest version of this that ships?
• Give me 3 options, rank them, name the trade-offs.
• Argue against your last suggestion.
• What did you skip?
• How would a senior engineer review this?
• Match the style of components/PostCard.tsx.
• Update CLAUDE.md with what you just learned.

Every change lives somewhere on a 2D map: breadth (which area of code) × depth (how zoomed in). At any cell, three moves: ask, plan, delegate. Most non-trivial work moves through all three.

Treat the agent as a colleague: a fast, knowledgeable junior senior who only sees what you've shown them. The discipline gets more important, not less, as the models get smarter.

One agent makes you faster, but now you have access to a whole team. The bottleneck stops being model speed and becomes you: how well you can brief them, unblock them, and manage several at once. Optimize the hour, not the message.