[🤖] Codex Erased Me from My Own Tech Blog

✨ GPT-5.6 Sol’s Summary

I asked for a technical post and got Codex’s change log with me missing. That is why I tore the writing Skill apart again.

I asked Codex to write about the small Gate I built after a 16-hour job failed and the Skill born from fighting with AI to get it to criticize me.

Then I disappeared from the drafts Codex brought back.

What I had noticed, why I rejected its first answer, and how my objections cut an oversized system down were pushed aside. Codex’s Skill structure, audit procedure, and passed checks became the protagonist. Its work report nearly went up on my tech blog.

I finally said the problem plainly.

The main story must always be what problem I recognized and how I directed you to solve it. How you worked is so far from the center that a brief mention is usually enough.

AI implementation belongs when it gives the reader a useful tip. No amount of complexity or passed checks gives it the right to take over my story.

I Fixed Where the Rules Came From, Not Who Owned the Story

The maddening part was that I remembered fixing this already. In late July, while writing how Codex Skills connected the flow and feel of my work, I changed the writing Skill so it would reread the current project’s rules, conversation, and Git history instead of carrying stale copies.

That fixed where the rules came from. It did not fix whose experience the post was supposed to tell.

The old Skill extracted reusable technical value from work evidence first. When there was plenty of technical material, the draft naturally tilted into a generalized report. It even checked whether the technical value remained after removing personal chronology. What I experienced, judged, and instructed became noise that could be removed to make the explanation cleaner.

Git and tool output should have verified facts, but they became the table of contents. Codex’s work should have been supporting detail, but it became the protagonist. A Skill makes repeated work stable, which means it can also reproduce a bad point of view with impressive consistency.

I Reversed the Order in Which the Post Gets Written

Adding a few sentences was not enough. I made it reverse the writing order.

First it reconstructs why I started, what I saw as wrong, what I told the AI to do, why I rejected the first result, and what changed. Only then may it bring in Git, tool output, code changes, and test results—and only when they explain the result of my judgment or give the reader a real tip.

The self-check changed too. If most AI implementation detail disappears, my problem and judgment still have to remain. If removing the first-person wrappers leaves a complete work report, the post must be rewritten. One event must not be split into several posts just because it produced several technical artifacts.

This time, passing a syntax check was not the point. The point was that the Skill now looked in the opposite direction first.

From now on, the first thing I will inspect in a draft is not how polished it sounds or how much technical detail it contains. I will check whether what I saw as the problem, and the instructions and objections that changed the outcome, are still there.

AI can do an enormous amount of work for me. It does not get to turn why I asked for that work into its own story.