You've fixed this four times
Same email. Same wrong detail. Fourth time this month.
You catch it, you fix it, the customer never knows. Total cost: about forty seconds — which is exactly why you've never asked why you keep paying it.
Forty seconds, four times a month, on one recurring task, is roughly half a workday a year — for a mistake you already identified and know the answer to. You aren't finding a problem; you're paying tuition on a lesson nobody wrote down.
Corrections go into the chat, and then nowhere
Watch where the fix lands. You edit the draft, send the good version, close the tab.
The draft got better. The system is exactly where it started. Next Thursday it produces the same output with the same flaw, because nothing about the correction outlived the window it happened in.
That's the asymmetry: your effort recurs, the benefit is disposable. Everything else in a business compounds; this resets weekly, quietly enough that nobody notices the treadmill.
A correction is worth making once
The fourth discipline in the four that decide whether AI is useful is loop engineering, and it's simpler than it sounds: a correction is only worth making once, and the loop is what makes sure it only happens once.
Almost nobody does it, because fixing the draft is faster in the moment and the cost only shows up months later, in a form nobody measures.
Capture, promote, verify
Capture. A running corrections log. One line per fix: what it got wrong, what right looks like. It has to happen in the moment, in under thirty seconds — if capturing takes longer than fixing, you'll abandon it inside a week.
Promote. Once a week, turn the log into edits to your standing context and playbooks. This is the actual work — about twenty minutes. You aren't fixing outputs; you're changing what the system knows, so that class of output stops appearing. Then clear the log.
Verify. Re-run the thing that failed — the same one, not something similar. If it still comes out wrong, your edit was too vague, which is the most common failure by a distance. "Be more careful with rates" changes nothing. "After-hours is $185/hr, after 6pm weekdays and all day Sunday" changes everything.
Your standing context is a snapshot
Your standing context and playbooks are a baseline — what you knew about your business the day you wrote them.
But a baseline is a photograph. Your prices moved. You dropped a service line. An account changed hands. Nothing about a well-written baseline updates itself.
The loop is the only thing that legitimately edits the baseline. Without one you don't have a system that learns your business — you have one that learned it once, in March, and has been drifting since. The corrections you make every week are that drift, as symptoms.
Six weeks of the Thursday recap
Every Thursday a recap goes out to your commercial accounts, and three things need fixing every week. It writes to the property management group in the chatty register it uses for everyone, when that contact wants clipped and formal. It leaves off the PO number, so AP kicks it back. And it keeps offering duct cleaning, which you stopped in the spring.
Week one, those three go in the log — ninety seconds. Friday they get promoted: that account's tone into the standing context, the PO number into the recap playbook's done-checklist, duct cleaning struck from the service list. Twenty minutes.
Week two, two are gone. The tone note was too vague, so you tighten it — no greeting, no pleasantries, bullets only — and verify against last week's email.
By week six the log is empty. Not because the AI got smarter, but because you stopped throwing the answers away.
What "gets better every day" honestly looks like
It isn't exponential. What happens is narrow: on the tasks you loop, corrections trend toward zero. On the tasks you don't, the output is exactly as good as day one, forever, however many model upgrades ship in the meantime.
Improvement isn't a property of the tool. It's a property of whether anyone is running a loop on it.
Something you do once a quarter isn't worth logging. The things with a cadence — the Thursday recap, the callback triage — are where six weeks of small edits turn into something that just works.
The loop is the asset
After six months, what you own isn't a subscription — it's a written, tested account of how your business actually does its recurring work. It survives a tool change, it survives the person who set it up leaving, and it's most of what makes a new hire productive in a week instead of a quarter. It's also why this is worth paying someone to run: setting up context is a project with an end, and the loop is a practice. For the playbooks it edits, see teach AI your playbook and standing context.
This week
- Open one file. Two columns: what it got wrong, what right looks like.
- For five days, write the line every time you fix an AI output. No exceptions.
- Friday, spend twenty minutes promoting the log into your context or playbook, then clear it.
- Re-run one thing that failed. Still wrong? Make the edit more specific and go again.
Six weeks on one task, then see how long the log is.
If what you need is someone running the loop every week so the system keeps up, that's the work. Book a 30-minute call and bring the correction you're tired of.