Somewhere in the last seven days you did something that took eleven minutes, that you have now done roughly forty times this year, and that would have gone perfectly well without you. Finding that thing is the whole job of this lesson.
Every lesson after this one teaches you how to build. This one decides what is worth building, and it is the one you might be tempted to skip, because it has no console in it and nothing lights up at the end. Skip it and you get a genuinely impressive agent that saves you about forty minutes a year. (You will know when it happens, because you will be showing the thing to someone and hear yourself say "and it does the whole job, apart from the part I still do.")
If you would rather be walked through this than fill in a table, open the companion prompt and paste it into a chat with Claude or ChatGPT. It runs these six steps with you one at a time and writes your sentence at the end. The exercise is identical either way, so take whichever gets you through it.
Write down everything you did in the last seven days apart from sitting with a client.
This is a dump rather than a schedule. Ten minutes, no editing, no sorting, no deciding whether something counts as work. The admin, the chasing, the rewriting, the "I'll just quickly," the thing you ended up doing at 9pm because there was nowhere else in the day for it to go.
These are the easy ones to forget:
| Column | The question | What goes in it |
|---|---|---|
| How often | How many times a month does this happen? | a number |
| How long | How many minutes each time, door to door | a number |
| Does it need you | Could a careful assistant who knows your practice do a first pass? | Yes, Partly, or No |
How often × how long ÷ 60 gives you hours a month, and that number is the prize. Forty minutes twice a week looks like nothing while it is happening and comes to nearly six hours a month, which is most of a client day.
The things you meant to do and never got to get scored as if you did them properly. How often it should happen, and how long it would take you. Recorded at what truthfully happened, which is zero times a month, they drop straight off the list in step 4, and those are frequently the best candidates on the page. Their cost shows up as clients you lost rather than hours you spent, so the arithmetic cannot see them at all unless you put the intended numbers in.
The third column decides everything, so it gets its own step.
An agent is good at gathering, watching for something, drafting a first version, summarizing, reminding, triaging, spotting a pattern, and having the thing you need sitting ready before you need it.
Where it falls down is deciding, judging a person, anything that runs through your license or your read of a room, and anything where being wrong costs you a relationship.
That looks like a clean split right up until you try to sort your own list with it, at which point almost nothing lands cleanly in either column. Most of your biggest drains are a mechanical ninety percent welded to a human ten percent. The follow-up message is mechanical all the way to the sentence that proves you remember who they are. Your weekly review has the same shape, mechanical the whole way through until the decision at the end.
So the question to ask of every item is what the last mile only you can do, and whether something else could carry it that far. That distinction is the whole lesson, and it is what keeps you from either automating nothing at all or automating the ten percent that was always yours.
Mark each item hand over fully, hand over the first ninety percent, or keep.
For everything you did not mark keep, multiply.
hours a month × (fully = 1.0 · first ninety percent = 0.7) = scoreSort descending. Your top three are the candidates.
What that number is actually worth, before you pick. Your score is a number of hours a month, and hours a month is easy to shrug at. Stretch it over a year and it stops being shruggable.
Ten minutes, every working day, is roughly forty hours a year. That is a working week you are handing back to yourself. The identical ten-minute job done once a month is two hours a year, which is an afternoon. Same task, same irritation, same feeling of dread when it appears, and one of them is worth a week of your life while the other is barely worth the setup.
So the thing that decides what goes first is volume, and it is almost never the thing that feels most annoying. The job that makes you sigh out loud is often monthly. The one costing you a working week is usually so routine you stopped noticing you do it, which is exactly why it never occurs to anyone to automate it.
The other half of the sum is what building costs you, and here that number is small and it keeps shrinking. Your first agent takes about two hours, and that is the expensive one, because you are learning the parts. Your second and third take roughly ten minutes each, since the box is already built and you are only writing a new job description into it.
Two hours against forty settles itself. Two hours against two is a genuine judgment call, so if your top candidate is something you touch once a month, look at the second one on your list before you commit.
Which of the three you build first is a judgment call, so work through these questions in order.
Then write one sentence and keep it somewhere you can find it again.
My first agent will ______________, every ______________, and leave me ______________.
Worked example, from a practice rather than a boardroom.
My first agent will find anyone who inquired and never got a reply, every morning, and leave me a draft message for each one.
If you are torn between two of them, drop it in the chat and we will talk it through. Both of them get built eventually, and the order matters less than you would think.
Some of what you listed should not be automated. It should stop.
A recurring task nobody needs is a thing to delete, and you have explicit permission to cross it off the list rather than build a small machine to keep it alive. Half the value of a time audit is the crossing off.
Then there is the trap on the other side. The follow-up that never happens and the review you keep skipping cost you clients rather than hours, so they never show up as time and never make the list, which is exactly why step 1 asked for the things you meant to do and never got to. Those are often the highest-value agent on the whole page.
And the one underneath both. You can spend more time building the thing than the thing was ever going to cost you. Somebody automates a ten-minute job over a weekend, it breaks, they fix it, it breaks differently, and six weeks later they are maintaining a small machine instead of doing ten minutes of work.
That failure is common enough to be a running joke among people who build software, and it is worth naming here because two things in this course keep you out of it.
You scored it before you built it. That is what step 4 was for, and it is why this lesson comes before any of the technical ones.
And what you are building is a job description, not machinery. When a script gets something wrong you repair a mechanism. When your agent gets something wrong you change a sentence and run it again. Fixing costs minutes here, which is what stops the maintenance spiral in that comic from being your spiral.
You finish with one sentence, in the format above. Lesson 5 names your first worker after it, and lesson 6 turns it into the job that worker does.