Every time I gave an AI a task, it came back wrong.
Not completely wrong — just slightly off. Wrong length. Wrong tone. Extra sections I never asked for. I'd revise, it would fix half and break something else. After three or four rounds, I'd wonder: is it the AI, or is it me?
Almost always, it was me.
Not because I wasn't clear about what I wanted. I was. The problem was I never told it what I didn't want.
AI fills in the blanks. That's the real issue.
When you say "write me an article," the AI assumes everything an article "should" have. You didn't say no headers — so it adds headers. You didn't say no summary — so it ends with three bullet takeaways you never asked for.
This isn't a capability failure. It's the AI doing things you didn't explicitly exclude.
I started noticing a pattern: every time a task went sideways, I had given detailed instructions on what to produce — and zero instructions on what to leave out.
The fix: make it show you the task breakdown before it starts.
I added one rule to my AI setup. Before executing any writing, organizing, or analysis task, the AI has to answer three questions and wait for my confirmation:
Goal — One sentence: what exactly are we doing?
Done criteria — How do I know it's finished? Specific and verifiable.
Out of Scope — What's explicitly excluded? At least one or two items.
It outputs this framework. I read it. If something's off — if the scope is too wide, if the goal is slightly misread — I catch it before the AI spends time on the wrong thing. I say "confirmed," and only then does it proceed.
This one change reduced off-script outputs by around 80%.
The exact prompt you can copy
Here's the instruction I added to my system prompt. You can drop this into any AI's custom instructions:
Before executing any writing, organizing, or analysis task, output the following structure and wait for my confirmation before starting:
Goal — One sentence describing what we're doing
Done criteria — Specific, verifiable conditions for completion
Out of Scope — Explicitly excluded items (at least 1–2)
Wait for me to say "confirmed" before proceeding. If you hit ambiguity mid-task, return to this framework and ask — don't improvise.
Principle: telling you what NOT to do matters more than repeating what to do. Out of Scope takes priority.
Who this is for
If you regularly delegate tasks to AI and end up revising the output, this is for you. Especially useful when the task itself isn't complex, but the result always has something slightly off.
The fix isn't to give longer instructions. It's to add one more constraint: what you don't want.
This template is based on my own setup. Adjust it to fit how you work — keep what's useful, drop what isn't.
I write to hold onto things. If this resonated, feel free to follow.