AI is a dance. 
One partner leads. 

The AI doesn't need to know how to move its pieces. It needs to know where you want to end up. When you tell it 'write me a function that does X', you've already failed, you're scripting the move instead of describing the position. Tell it what the code should do to the user, or how the reader should feel, or what state the system should be in when you're done. The AI will find the steps. You're the one with taste.

Amateurs describe implementations. Leaders describe outcomes. The harder the outcome is to put into words, the more important it is that you try, because the reason you're talking to AI is exactly to get help with the stuff you can't reduce to a mechanical description. If you can describe it mechanically, you can write it yourself.

The biggest single failure mode in AI work is assuming the model knows what you know. It doesn't. It knows general programming, but it doesn't know your repo, your domain language, your team conventions, or the three bad decisions you already made last Tuesday. All of that has to be in the prompt, not around it.

I pay the context cost up front. When I start a session I dump: the repo structure, the relevant files, the constraints I'm working under, and the vocabulary I use. It feels wasteful the first time, you're writing paragraphs the AI might not use. It isn't wasteful. The AI can't ask clarifying questions the way a colleague would. Either you tell it up front, or you correct for an hour.

When the AI drifts, don't re-explain what you wanted. Name the failure and tell it to invert. 'You're overexplaining' beats 'please be more concise.' 'You're guessing, stop' beats 'please verify before answering.' 'This is boring' beats 'please write something more compelling.'

Why it works: the AI already knows what good looks like along hundreds of dimensions. It just chose a bad mode this time. Your job is to tell it which axis it's wrong on, not to re-describe the whole target. The correction is shorter, the re-anchor is faster, and you don't lose the thread of the session.

One of the highest-leverage moves is asking the AI to talk you out of your own idea. 'I'm thinking of doing X. What breaks?' 'Is this a good idea? Give me the real answer, not the polite one.' 'Challenge this.' The AI is a resistance tool, you can stress-test your own thinking at the speed of typing.

Most people never use the follower this way because they're afraid of being wrong in public. There's no audience. You're talking to a machine. Being wrong to the AI costs nothing. Being wrong in production costs everything. Use the cheap resistance.

Don't wait for the perfect prompt. Don't wait for the perfect plan. Ship rough, look at the output, correct. The AI can polish what exists ten times faster than it can generate it from scratch, but only if you give it something to react to.

I push commits while they're still rough. I correct in the next message. I don't apologize for the first draft; I don't defend the second. Most of the value is in the third draft, and you can't get to the third draft without writing the first two fast.

Every extra word in your prompt dilutes the signal. 'Please could you help me write a function that…' is noise. 'Write a function that…' is half the length and has more information density. Fragment over sentence. Verb over noun. Direct over polite.

This isn't about being rude. The AI doesn't care about politeness, politeness costs tokens and buys nothing. Save your articulation for the content of the ask, not the form of it. The buyer of your output is the user of your product, not the model.

A prompt template that worked last week won't work next week when the model ships a new version. A principle you actually understand will work on any model you'll ever use, including the ones that don't exist yet. Invest in the principle, not the recipe.

Every time I write a new prompt, I'm trying to make it shorter next time by writing the principle behind it down somewhere I'll find again. The recipes are ephemeral. The principles compound.

If you tell the AI something once and have to tell it again tomorrow, you're trading away the biggest compounding asset in AI work. The follower is only as good as what you've taught it. Every lesson you don't persist is a lesson you'll re-teach at the cost of time and patience.

I built an entire memory system, files that every conversation loads first, so future sessions start with everything I already taught it. The first lesson is expensive. The hundredth is free. Most people pay the first-lesson price every session. That's why their output doesn't improve, they're still teaching the same follower the same steps every morning.