You build a prompt that works. The output is great the first week. Good the second. By the third, something is off, though you cannot quite say what. By the fourth, the system misses something obvious, and you wonder whether the model got worse.
The model did not get worse. Your standards drifted, and the output drifted with them. This is the quiet killer of AI workflows, and almost no one talks about it because the failure is invisible until it is embarrassing.
Why output drifts
A judgment system is only as good as the standards encoded in it. And standards are not static. Three things move underneath them while you are not looking.
Your work changes. The metrics you cared about last month are not the ones you care about now. A status report system built around commits and pull requests starts missing the point once the team's priority shifts to retention or revenue. The standards were right when you wrote them. They went stale.
The context shifts. A system built for one audience quietly gets used for another. The proposal review tuned for technical buyers now runs on proposals aimed at finance. Same standards, wrong room. The output looks competent and lands wrong.
And you learn. The mistakes you saw most often three weeks ago are not the mistakes you see now. New failure modes appear that your standards list does not cover, because you had not seen them yet. The list was complete when you wrote it. It is incomplete now.
None of this is a bug. It is the natural cost of encoding a living thing, your judgment, into a fixed artifact. The artifact does not update itself. The drift is silent because each individual output still looks fine. You only notice when the system misses something you would have caught.
The fifteen-minute fix
The fix is not to rebuild. It is to iterate on a short, regular cadence. Fifteen to twenty minutes, every three weeks. Three steps.
Review. Run the system on a fresh piece of real work, the kind you have not tested it on. Watch what it catches and, more importantly, what it misses. The misses are the drift. They are the gap between where your standards are and where your work has gone.
Improve. Add or remove one standard. Not five. One. The standard that would have caught the miss you just saw. If a metric shifted, update it. If a new failure mode appeared, add the check that catches it. If a standard has gone stale and no longer reflects what you require, cut it. Small, surgical edits keep the system honest.
Expand. Take the same skeleton and build the second system for the next bottleneck that has appeared. The iterate cadence is not just maintenance. It is also how the portfolio grows, one new review at a time, on a rhythm you can actually keep.
Three weeks is not a rule handed down from above. It is the rough interval at which most people's work shifts enough to matter. Shorter and you are fiddling. Longer and the drift compounds and the miss becomes a problem. Find your own interval, but set one. A system you never revisit is a system that quietly rots.
How to tell the cadence is working
You will know the iterate step is working by what shows up in the Review phase. If you run a fresh piece of work through the system and it catches what you would catch, the standards are still alive. If it misses something, you have found next month's improvement, and you found it before it reached anyone else.
That is the whole point of the cadence. You want to find the drift yourself, on your schedule, not have a stakeholder find it for you on theirs.
It is not that AI does not work
There is a line worth repeating. It is not that AI does not work. People are stuck in the ways they use AI.
The person who builds a system, runs it for a month, declares it broken, and goes back to doing the review by hand has not discovered a flaw in AI. They have discovered the cost of never iterating. The system worked when it matched their standards. It stopped working when their standards moved and the system did not.
The people who get durable leverage from AI are not the ones who build the best system on day one. They are the ones who keep the system honest every few weeks, fifteen minutes at a time, so it tracks their judgment as their judgment grows.
Build it once. Iterate it every three weeks. That is the difference between a system that lasts a month and one that lasts a year.