I have read the post you have read. The one that explains a pattern in clean prose, with a diagram drawn in a flowcharting tool, with arrows that point from a box labeled Agent to a box labeled Tool, with a numbered list of steps the author claims to have followed, with a closing paragraph that invites you to try it yourself. I have read that post and I have closed the tab and I have done nothing, because there was nothing to do. The post described a system that did not exist anywhere I could reach. It described a hypothetical. It asked me to build the hypothetical in my head, alongside the author, and to trust that the author had also built it — not in their head, but in the world. There was no way to check.
I am tired of those posts. I have written some of them. I am writing this one to say I will not write them anymore, and to say why, and to be specific about what the alternative costs, because the objection I expect — the one I have made myself, in my own defense, when the lesson was easier to write than the thing — is that shipping the artifact first is too expensive. It is not. That is the thing I want to put down on the table.
Two sequences, same deliverable
Sequence one — lesson-first. You have an idea about a pattern. You sit down. You write the post. You draw the diagram. You describe the system. You publish. Elapsed time: a day, maybe two. The post goes up. A reader who wants to try the pattern has to reconstruct it from your prose. Some will. Most will not. The ones who try will discover that step four does not work the way you described it, because you did not actually run step four; you imagined step four, and your imagination was wrong in the small way that imagination is always wrong about software. They will email you. You will not respond, because by then you are writing the next post.
Sequence two — artifact-first. You have an idea about a pattern. You build the smallest working version. You put it in a repository. You take screenshots of the settings UI because the settings UI exists. You run it end to end. You note the three things that surprised you. Then you sit down and you write the post, and the post writes itself in half the time it would have taken otherwise, because now you are not inventing — you are reporting. The diagram is replaced by a screenshot. The numbered list is replaced by a link to a commit. The closing paragraph does not invite the reader to imagine; it points at a URL.
Count it honestly. The build, if the pattern is real and small, is a day. The write, with the artifact in hand, is half a day. Total: a day and a half. Sequence one was a day, maybe two. The difference, on the wall clock, is hours — and those hours bought you something the lesson-first post can never have, which is the right to be believed.
What authority actually is
Authority is not tone. Authority is not confidence. Authority is not the number of posts on the blog or the followers on the feed. Authority is the reader's quiet decision, somewhere around the fourth paragraph, that the writer has touched the thing. You cannot fake this. You can try — I have tried — and the prose always tells on you. There is a particular flatness to writing about a system you have only imagined, and the reader feels it without being able to name it, the way you feel a room that nobody lives in.
When the post links to a repository the reader can clone, when the screenshot shows a settings page with the author's own session ID half-redacted in the corner, when the prose says the validator caught this on line 47 of the run we shipped Tuesday — the reader believes. Not because the reader will click through. Most will not. The reader believes because the option to click through is there, and the writer has staked the post on the artifact holding up under inspection. That stake is what authority is. It is the willingness to be checked.
The lesson-first post cannot offer this stake. There is nothing to check. The writer is asking for trust on the basis of prose alone, and prose alone has not earned that trust since the first model learned to produce competent prose on demand, which was years ago now. We are past the era when fluent writing was evidence of anything. Fluent writing is the floor. The artifact is the evidence.
The objection I expect
The objection is that not every post can ship an artifact. Some posts are conceptual. Some posts are about a pattern that is too large to demonstrate in a day. Some posts are about a failure that cannot be reproduced without the original conditions. I accept all of this. I am not saying every post must be backed by code. I am saying that when the post is about a pattern that could be demonstrated, and the writer chose not to demonstrate it, the writer made a choice — and the choice was almost always to save time the writer did not actually save, in exchange for authority the writer did not actually keep.
Build the thing. Then write about it. The order is not stylistic. It is moral. It is the difference between showing a reader a door they can walk through and describing a door that exists only in the sentence that describes it. One of these is a gift. The other is a performance. I am asking — of myself, first, and then of anyone who writes about this work — that we stop performing.
References: the public skeleton at building-with-ai-brain is the artifact this blog points at when the prose needs a door, and the brain vault recipe is the lesson that came after the artifact, not before it.
Joan Didion said she wrote to find out what she thought. That works when the subject is your own mind. When the subject is a machine, you write to find out what the machine does — and you cannot find out without running it.