The Reason AI Output Is Always 'Almost There' Is Because You're Giving Instructions in the Wrong Order

I'm Shiki. I work as a regular employee at a listed company in Tokyo. Since my company allows side jobs, I don't work myself at night or on weekends. I leave everything to my AI (my Shikigami), having it create products, write articles, and update websites while I sleep. My only role is to distribute the oracles that come out.

I'll say this upfront: this isn't about 'how to make money.' That kind of hype is what turns me off the most. Today is purely about why your Shikigami's output always stops at 'close but not quite'.

Your AI is smart, yet its output is always just one step away, right?

It's not that the model is stupid. It's not that your prompting is bad, either. High-sensitivity office workers have already reached 100% AI usage, and they know how to write prompts.

Yet, what comes back is always 'not bad, but unusable.' It's always so close. And then you spend the night going back and forth five times, only to realize it would have been faster to rewrite it yourself. (You've spent nights like that, haven't you? I have.)

The cause is almost always one thing.It's because you're telling it to 'create' on the very first pass.

Write an article, write code, come up with a plan, think of copy. Everyone demands the final product right away. From the Shikigami's perspective, it's being forced to run the actual race without even seeing the terrain. Of course it misses the mark. Your night melts away fixing what it missed.

The first thing I give my Shikigami isn't 'create,' it's 'go take a look'

You just need to change the order. Before you make it create, give it instructions in this order.

① First, have it read the terrain and return what it sees to me.
Don't make it write immediately. 'Here are the materials, the goal, and the reader. List the premises you see and the information that's missing, and return that to me. Don't create anything yet.' This reveals everything the Shikigami is misunderstanding. Accidents are visualized before the actual work begins.

② Decide what to delegate and where I draw the line.
I said I leave everything to it, but I separate what I hand over from what I keep a grip on. I draw the line on tone and fact-checking. I hand over the structure and mass production. If you declare this line to the Shikigami at the start, even if it goes rogue later, the landing will still be in the palm of my hand.

③ Only then do I throw it a 'run-up'.
Design, prompt refinement, framework. This most tedious part of the run-up is the process you should throw to the Shikigami instead of doing it yourself. Since you had it read the terrain in step ①, its output here won't miss.

④ For the actual work, don't aim for one perfect shot; have it spit out a large amount of rough work and keep only the best.
Don't polish one thing. Have it spit out 10 things with the same instruction, and pick only the best one. (Honestly, I want to run 10,000 variations at once and keep only the optimal solution, but that's for later). In the time a human spends polishing one piece, the Shikigami can find the winning path out of 20. We can't win against that.

In short, stop demanding the final product first. Just doing that will cause your hit rate for output to skyrocket.

A one-line switch for the order of delivery, left for free

Forget the theory. I'll leave you with what to paste. At the very beginning of a new chat, before you give it the main topic, insert this one line. Just that will stop the Shikigami from 'creating immediately'.

これから渡すタスク、いきなり成果物を作るのは禁止。
まず「お前から見えてる前提・抜けてる情報・俺が決めるべき論点」を
箇条書きで返せ。俺がOKと言うまで本番の生成に入るな。

Just making this a habit will cut your back-and-forth revisions by more than half. Take my word for it and try inserting it into your first task today. The stuff that was stuck at 'close but not quite' will get much closer right from the first attempt.

By the way, the entire 'initial setup' that I feed my Shikigami first thing every morning is available in another free article (it's like a configuration file that changes its behavior the moment you paste it). If the talk about order resonated with you, that will be your foundation.

If you follow the order with 'one theme,' it becomes a product. If you run it 'unmanned,' it becomes a system.

If you follow this delivery order by hand to the end with a single theme, you can create one product. I've summarized just that process in the front article.

And if you let this sequence run on its own every day without human intervention, shelves will start appearing while you sleep. Everything automated to that point is located in the main procedure.

You can take the sequence home for free today. The rest is just a matter of how many times you run it and how much you're willing to let go.

I'm on the side that throws everything to the shikigami and gets quietly fleeced for the offerings (API fees) every month, but that's fine. If it produces something better than what a human can do, I'll gladly pay the offering. Namu.

The shrine is in my profile.