Yesterday my own publishing pipeline refused to publish one of my posts. It exited with an error and told me the link in it had no tracking on it, so any traffic it sent would arrive unattributable. The fix took about four minutes. The alternative took about four seconds, and I knew exactly what it was.
There is an environment variable in my system called QC_ALLOW_WARN. Setting it to 1 tells the quality gate that I have seen the warning and want to proceed anyway. It exists for good reasons. It is also the single easiest thing in my entire operation to do wrong, because nobody is watching me do it.
This is the part of working alone that I did not expect to be hard.
What I thought the hard part would be
Before I did this full time, I assumed the difficulty of working alone would be loneliness, or motivation, or the silence where colleagues used to be. Those are real, and they are manageable. You can schedule around them.
The thing I did not anticipate is that I am the only person who will ever see the shortcut. Not the only person who could stop me. The only person who will ever know it was available.
In a company, a gate that fires in front of you has an audience. Someone reviews the pull request. Someone signs the thing off. The override is visible, and its visibility is doing most of the work, more than the gate itself is. Alone, the override is a keystroke in a terminal that nobody will ever read, followed by a green result that looks identical to the one you would have got by doing it properly.
The output is the same either way. That is the whole problem.
I know it is a problem because I have already done it
I am not writing this as someone who resisted temptation. I am writing it as someone who checked his own record and did not like it.
At the start of this month, two consecutive runs published with that override set. Both times the reason was defensible. The gate was flagging currency figures inside compliance posts, all of them statutory thresholds quoted from the tax authority, and it flagged sixteen of them without being right about a single one. So I overrode it. Twice. The content was fine. The judgement was fine.
The habit was not fine, and I only saw that when I went back and counted. Sixteen warnings, zero real problems, two overrides. What I had actually built, without intending to, was a gate I had learned to wave through. A gate you wave through by reflex cannot stop the thing it exists to stop, because by the time a real problem arrives you are no longer reading the warning, you are dismissing it.
I wrote about a related failure in the watcher I built that said nothing. This is the mirror image. That system was silent when it should have spoken. This one spoke, correctly, and I had trained myself not to listen.
What I changed, and what it cost
The fix was not more discipline. Discipline is what I had been relying on, and it had produced two overrides in two days.
I rebuilt the gate so it distinguishes between three cases instead of one. A figure with a statutory source next to it is now suppressed silently and counted, because it was never a real risk. A figure with no signal either way still warns. And a figure that reads like our own price list is now blocked outright, in a way the override cannot touch at all.
That last part is the bit that matters. I removed my own ability to wave through the specific thing I most need to be stopped from publishing. Not because I expect to try, but because the version of me that would try is tired, behind, and not reading carefully, and that version deserves a locked door rather than a warning.
Yesterday's refusal was the new arrangement working. The link had no tracking parameter, so the gate stopped the post. I could have overridden it, and the post would have gone out looking completely normal. Instead I added the tracking, verified the destination actually loaded, and then widened the gate so it would catch the same defect on a domain it had not been covering. Four minutes. Nothing published that day used the override.
I am aware of how small that is. Four minutes and a tracking parameter. It is not a dramatic story. That is rather the point: none of these decisions are dramatic while you are making them, which is why the record of them is the only reliable evidence of what you actually do.
The part that is genuinely hard
It is not the discipline. It is that there is no external signal telling you whether you are getting it right.
If I cut a corner, the post still publishes. The site still loads. Nobody emails me. The consequence, when it eventually arrives, does not arrive labelled, and it does not arrive on the day I earned it. It shows up months later as a number that was wrong, or a link that never got attributed, or a claim I cannot source when someone finally asks.
So the only feedback available is the kind I have to build for myself and then choose to read. That means my standards cannot live in my intentions, because my intentions are the thing under pressure. They have to live in something that will refuse me. A gate that exits non-zero. A check that runs whether or not I remember it. A rule that I wrote on a day when I was thinking clearly, enforced on a day when I am not.
Working alone means being both the person who sets the standard and the person the standard is set against, and those two people are never in the room at the same time. The first one is calm and has time. The second one is behind schedule at six in the morning and has a green button available.
Most of what I have built this year, when I look at it honestly, is the first one leaving instructions for the second one.
Frequently Asked Questions
Why not just remove the override entirely?
Because some warnings genuinely are false, and a gate with no escape hatch gets deleted the first time it blocks something urgent and wrong. The answer is not zero overrides. It is that the highest-risk category cannot be overridden, and every override that does happen is recorded.
How do you know whether you are overriding too often?
By counting, after the fact, and looking at the ratio rather than any single decision. Sixteen warnings and no real problems told me the rule was miscalibrated. Two overrides in two days told me I had adapted to it rather than fixing it.
Does this apply to anyone who is not a solo operator?
I think it applies to anyone whose work is not routinely seen by someone else before it ships, which is a much larger group than people who work alone. The relevant variable is not headcount, it is whether the shortcut has a witness.
Where this connects
Today I published something on the business site about what actually governs AI use in South Africa right now, which grew out of a government policy that was withdrawn because nobody verified the sources in it before it went out.
I wrote that piece and then realised I had spent the previous day doing exactly the small version of the same thing: standing in front of a check, with the ability to skip it, deciding whether to.
The government's version had a whole department in it and it still happened. Mine has one person. I do not think that makes me safer. I think it makes the written-down rules the only thing between the two outcomes.