AI as a Partner, Not a Tool

For a long stretch I treated AI as a tool that types fast: I already knew what I wanted, and it implemented it. That is not worthless, but the output always stopped at the line marked “what I could have thought of anyway,” and never went a step further.

The real change came when I changed the method: instead of issuing instructions, I started by writing a design brief together with it.

Where the line between tool and partner falls

Using it as a tool means: I know the answer, it works the keyboard. Using it as a partner means: I do not fully know the answer, so it has to help me sharpen the question first — and it is entitled to tell me that what I am asking for contradicts itself.

The dividing line is not how strong the model is; it is whether I give it a chance to disagree with me. And for disagreement to be possible, there has to be something both of us can point at — not a sentence buried in a chat log, but a file.

The design brief is a joint artifact

My design briefs run on a few hard rules:

• Never hand me “the best option.” For any visual direction, produce three real, viewable drafts whose styles genuinely diverge, and let me choose. A single carefully argued proposal has quietly taken the choice away before I arrived.
• Decisions become records. Context, decision, consequences — three parts. Why a decision was overturned is worth more than the decision itself; two weeks later it saves you a mountain of repeated debate.
• Disagreements stay in the file. Where it argued against me and I did not take the advice, that gets recorded too, along with my reasoning. Then the next window to pick this up does not fall into the same hole twice.

A concrete example

This site’s palette was settled exactly that way.

I had a direction in mind when I started. Per the rule, it first gave me three homepage drafts in genuinely different styles — and the one I ended up choosing was not the one I had walked in wanting.

The more interesting round came later. After launch I suddenly recalled that my main site seemed to use a different palette, and asked whether that one would suit better. It did not simply agree with me. It went and read the source of several branches of the main site, then came back to report: every branch shares one and the same palette, and the “other color” I remembered was its dark mode.

That fact changed the decision — we ended up adopting the main site’s entire color system, dark room included. Had I been using the AI as a tool, it would most likely have cheerfully implemented a palette that did not exist, faithful to my memory.

Speed is not the point

This site went from empty repository to a live homepage in two days. But the two days are not the point, and are even a little misleading — what actually remains is not the site, it is the design briefs and decision records the process left behind. They outlast the pages by a wide margin: when the next project starts, we do not begin from a blank sheet, we begin from a stack of architectural decisions already discussed and settled.

This kind of partnership does have a price, of course: you have to write your ideas down so they can be argued with, and you have to accept conclusions like “I remembered it wrong.” What you get in return is worth it — the finished work turns out a little better than anything you would have reached alone.