Writer Notes That Leaked Onto Live Pages

What was broken

Every day a posting sheet tells us what goes on which website. That sheet was written for a person. Each website section ended with a few lines meant only for her. One line listed related reading. One line was labelled as the call to action. One line gave the character count.

She knew these were instructions. She read them and did not paste them. So they never reached a reader.

On 16 Sep 2026 I moved website publishing from that person to a script. The script copies each section as it is. It had four checks before anything went live. One for the title. One for empty fields. One for who owns the work. One for names that should not appear. None of the four looked for writer notes. Nobody had asked it to, because a human had always done that filtering without being told.

On 20 Sep 2026 a blog post went live with all three note lines showing at the bottom. A reader could see the character count.

The same day, five older posts were cleaned up by hand. That fixed the pages. It did not fix the cause. No check was added. So the problem was free to come back on the next post.

What I built

I built a small filter that runs before every publish. It has two layers.

The first layer removes the note shapes I already know. The related reading line goes. The character count goes. Topic cross references go. For the call to action, only the label goes. The sentence itself stays, because that part was written for the reader.

The second layer looks at whatever is left. If anything still looks like a note, the publish stops. It does not guess. It does not let it through. A new kind of note I have never seen stops the lane instead of going live.

That second layer is the part that matters most. The first layer only knows yesterday's mistakes. The second one catches tomorrow's.

The filter is now the fifth check in the publish run, next to the other four. The publish run also sweeps the live site for notes that got through before the check existed.

What changed

On 24 Sep 2026 I ran a fix across the live site. It cleaned 3 pages in one go.

Then I ran the publisher again on the sheet that had caused the leak. The log shows the filter removed 3 note lines from the post before anything else happened. The call to action line lost its label and kept its sentence. The new notes check passed. The post was already live, so the run skipped it and overwrote nothing.

That day the live-site sweep ran 7 times, across dry runs and real runs. All 7 came back clean.

From the leak on 20 Sep to a lasting fix on 24 Sep took 4 days. The first fix, on 20 Sep, was by hand and did not hold. The second fix was a check, and it does.

What I would do differently

When I took a job off a person, I copied what she did. I did not copy what she knew. Skipping those note lines was never written down anywhere. It lived in her head. The script could not know it.

Next time I hand any task from a person to a system, I will sit with the person first. I will ask one question. What do you look at and quietly ignore? Those silent rules are the ones that break first.

I would also never accept a hand clean-up as the end of a fix. Cleaning five pages felt like progress. It was not. If a fix does not add a check, the same mistake is just waiting for the next run.

And I would build the "stop if unsure" layer first. A list of known bad shapes is useful. A rule that blocks anything unknown is what keeps the site safe.