Quick confession before we start. In my first month with AI, I was sure it just couldn’t help with anything I’d want to keep.
I totally blamed the AI.
Turns out it was my prompt.
I was asking terrible questions
I went back through my old chats recently, and it was a little embarrassing.
The same eight mistakes, over and over. Every one of them comes down to a missing piece of information, and when the AI has to guess, it guesses the most ordinary thing it can.
So this week, no single trick.
Instead, all eight mistakes, what goes wrong, and what I ask now.
To keep it concrete, every example uses the same scene: Marta, at a bus stop in the rain, the day after her mother’s funeral. Her sister Clara didn’t come. The bus arrives, and Marta doesn’t get on.
1. “Make it better”
The problem: “better” means nothing to the AI. Shorter? Sadder? Faster? So it picks its default, which is usually more adjectives and more drama.
The fix: name the one thing you want to improve, and ask for a diagnosis before any change.
Me: The tension in this scene drops somewhere in the middle. Point to the exact sentence where you’d stop caring, and tell me why in two lines. Don’t rewrite anything.
AI: At “She thought about calling Clara.” Up to there Marta acts. Here she only thinks, and the thought tells us what the rain already showed.
2. No context
The problem: without genre, point of view or reader, the AI writes for an average reader of an average book.
The fix: three lines before every request. What this is, who it’s for, which draft you’re on.
Me: This is a scene from a literary novel, close third person on Marta, second draft. Readers already know Clara skipped the funeral. What in this scene is explaining something they already know?
AI: Paragraph two: “Of course Clara hadn’t come.” The reader knows. The line sounds like you reminding yourself. Her silence at the bus stop could carry it instead.
3. Taking the first answer
The problem: the first answer is the most likely one, which makes it the most average one.
The fix: ask for several options, and make each one less obvious than the last.
Me: Give me six reasons Marta might not get on the bus. Make each one less obvious than the one before.
AI: 1. She’s too angry to sit among strangers. [...] 6. She’s realized nobody is waiting for her anywhere, and getting on the bus would make that official.
Number one I’d have written myself. Number six I wouldn’t have found alone.
4. Asking for text instead of questions
The problem: when the AI writes the dialogue, it writes it in its voice, not yours. Then you spend an hour un-AI-ing it.
The fix: ask it to question you instead. (Remember Elena from issue #1? Give the AI a role, not a task.)
Me: Ask me four questions about what Marta and Clara want from each other when they finally talk.
AI: 1. What does Marta want Clara to say, and what would she do if Clara actually said it? [...] 4. Is there something Clara knows about their mother that Marta doesn’t?
That last one opened a whole subplot I hadn’t planned.
5. “Do you like it?”
The problem: AI assistants are built to be agreeable. Ask if it likes your scene and it will. Of course it will.
The fix: give it a specific, unimpressed reader, and ask for weaknesses.
Me: You’re a tired editor reading this at the end of a long day. Name the three places where you’d stop reading. Don’t be unnecessarily kind.
AI: 1. The opening line about the rain. I’ve read it a thousand times. 2. The paragraph about her mother’s coat, which tells me she’s sad twice. 3. The ending, which stops a beat before anything happens.
6. Pasting the whole chapter
The problem: with forty pages, the feedback spreads thin and turns vague. “The pacing could be tightened in places.”
The fix: one scene at a time, with one question.
Me: Here’s only the bus stop scene, about 400 words. One question: does the reader feel that Marta made a decision, or just that she missed a bus?
AI: Mostly that she missed a bus. She “watches it pull away”, which is passive. One small physical choice, like stepping back from the curb, would make the decision visible.
7. Trusting the facts
The problem: the AI states made-up details with total confidence. A bus route, a funeral custom, how long a medicine takes to work.
The fix: ask it to flag what needs checking, then check it yourself.
Me: List every fact in this scene a reader could check: places, dates, how things work. Mark the ones you’re not sure about. Don’t correct them.
AI: 1. “The 47 to Porta Romana”: I can’t confirm this route exists. 2. A funeral the morning after a death: customs vary by country and religion, worth checking for your setting.
8. Not saying what to leave alone
The problem: the AI edits everything it touches. Ask for help with pacing and your dialogue comes back rewritten too.
The fix: say what’s off-limits, and ask for notes instead of a new version.
Me: Help me with pacing only. Don’t change any dialogue, don’t add events, and leave the last line exactly as it is. Put your suggestions in brackets instead of rewriting.
AI: [Paragraph one has “slowly” and “carefully” before anything happens. Cut one.] [Bring the bus in one sentence earlier: right now the wait lasts longer than the anger.]
Your turn (10 minutes, I promise)
Pick the mistake from the list you recognize the most. Be honest!
Take a scene or section you’re working on and run the “Me” line from that example, with your own details swapped in.
Read the answer and underline the one note you didn’t see coming.
Rework 100 words of your scene around that note. In your own words, please.
Pay attention to the mistake you were sure you never make. In my experience, that’s the one hiding in your chat history.
A few last thoughts
Look at what all eight fixes have in common: none of them asks the AI to write the scene. They ask for a diagnosis, options, questions, a tough reader, a list of things to check.
The words on the page are still yours.
If you write non-fiction, I guess mistake seven is the one that bites hardest. Run that example on every chapter, and check every flagged fact by hand.
And be patient with your own first month.
Mine was messy, and I’m glad I didn’t close the tab at the first hiccup.
Which of the eight is yours? Hit reply and tell me your number. And if you’ve made a ninth mistake I missed, I want to hear about that one too.
Until next week, Wilma