Automation is where AI stops being interesting and starts being useful. It is also where most people lose a weekend building something that quietly produces nonsense.
Start from a prompt you already trust
If you have not run the prompt manually five or six times without editing the output, you are automating a guess. Manual repetition is the test suite.
Do the arithmetic first
Three questions, before any tool is opened:
- How often does this actually run? Weekly is a workflow. Twice a year is a hobby.
- How long does each run save? Under roughly thirty minutes a month total, and you are building for fun.
- What happens if it produces something wrong at 3am, and who finds out?
That third question decides how much checking you build in.
Write the steps down before building
Numbered steps, in plain language, marking which ones still need a human:
- Trigger — what starts it.
- Fetch — what material gets gathered, from where.
- Prompt — the exact text, with the material inserted.
- Check — what makes an output acceptable.
- Deliver — where the result lands, and who sees it.
Most automations that go wrong are missing step 4.
Keep the path fixed
Fixed-path tools — Zapier, Make, n8n — do this well precisely because they do not improvise. You can see which step failed and what each run costs. Save the self-directing agents for work that genuinely varies every time.
Ship it half-automated
The best first version does everything except send. A draft lands in your inbox, you glance at it, you press send. After a few weeks you will know whether the checking step can be trusted to run without you.
Try this today
Write your working prompt out as the five numbered steps above, and mark the one step you are not yet willing to let run unattended.