August 4, 2026
Do It Twice, Then Write It Down as a System
Once is a task. Twice is a pattern. The third time you are about to do the same thing again, stop and write the steps down instead. That is the whole rule for knowing when something is ready to become a system.
Most people systemise too early, on something they have done once and will never do again, or far too late, on the thing they have quietly repeated for two years.
Why twice is the number
Once tells you nothing. Plenty of work happens once and never returns, and building for it is time you do not get back.
Twice tells you it has a shape. You now know which parts stayed the same and which changed, which is exactly the information a system needs and the thing you cannot know from a single run.
The three things you need first
None of them is software, and this is where most projects skip ahead and pay for it.
- What triggers it. The moment the work should start, without anyone remembering to start it.
- What the steps are. Written in order, before you decide who runs them.
- Who checks it. The person or agent that confirms the result before it counts as done.
Decide who runs it last
Agent or person does not matter yet, and deciding it first is how you end up building an agent for something a checklist would have fixed.
Get the steps right. Then choose what runs them. That order is also what makes the choice easy, because a written-down process makes it obvious which parts need judgment and which never vary.
And the thing that makes any of it visible while it is running is asking your AI to show its work.
When should I turn a task into a system?
When you have done it twice. Once is a task, twice is a pattern, and the third time is the moment to write the steps down instead of doing it again.
What do I need before automating a process?
Three things, none of them software: what triggers the work, what the steps actually are, and who checks the result before it counts as done.
Should an agent or a person run a process?
Decide that last. Get the steps written down first, which makes it obvious which parts need judgment and which never vary.
Why not systemise something after doing it once?
One run tells you nothing about what stays the same and what changes. Two runs show you the shape, which is the information a system actually needs.