There is a note in my phone, timestamped March 2019, that reads: 'Forensic accountant finds out her dad embezzled from the same charity her daughter now runs. Thriller.' That note has survived two phone upgrades and one cloud migration. What it has not survived into is an actual manuscript, because for six years I could not figure out where the book starts. The office? The discovery? The daughter's gala? Every Saturday I'd open a blank document, type 'Chapter One', stare at it like it owed me money, and go reorganize the spice rack instead.
I am not a person who lacks ideas. I am a person who is terrified of the first sentence, because the first sentence implies all the sentences after it, and that's the whole novel, and the whole novel is too big to look at directly without flinching. So I avoided it, productively, for six years. I told people I was 'percolating.' What I was doing was the literary equivalent of standing at the edge of a cold pool for half a decade, narrating to myself about water temperature.
What the prompt actually did differently
I'd tried generic 'write me a scene' prompts before and gotten generic mush — a man in a trench coat, rain, a cigarette, all the noir furniture and none of the furniture being mine. The thing I like about Turn a One-Sentence Logline Into a Cold-Open Scene is that it forces specificity into the inputs before the model ever writes a word. It doesn't ask for 'a thriller.' It asks for genre, POV character, scene moment, and a setting hint, and that fourth field is where it earns its keep.
I gave it my logline, set genre to 'literary thriller', POV to the accountant, scene moment to 'the second she finds the discrepancy in the old audit', and for the setting hint I just typed 'somewhere hot and ugly.' I handed it to Claude Opus 4.8 because Opus is the model I trust to hold a tone and not slide into thriller cliche, and that instinct paid off. The XML-tagged structure of the prompt clearly does something to how it behaves, too — there's an explicit avoid block that bans waking-up openings and weather, and for once a model actually obeyed it, which after years of generic outputs felt borderline emotional.
The detail I didn't write but kept
What came back was 520 words set in a self-storage unit in Phoenix in July. My character is kneeling on a concrete floor, a banker's box open in front of her, and her reading glasses keep fogging up because the unit has no air conditioning and she's leaning over the box breathing onto her own lenses. She has to keep taking them off to see clearly. I did not write that. I would never have thought of Phoenix, or the storage unit, or the glasses.
And the glasses-fogging thing — a woman whose entire job is seeing clearly, who literally cannot see the document in front of her — became the spine of my whole first chapter. It's not subtle, but it's the right kind of unsubtle, the kind a reader feels before they notice it. I found it in the output and recognized it as true. That's the part nobody tells you about writing with a model: a lot of the work is recognition, not generation. The model produces twenty details; nineteen are filler and one is load-bearing, and your only job — the job that's still entirely human — is to know which one.
I want to be clear about what I did NOT keep, because that matters just as much. The dialogue in the original output was stiff and a little on-the-nose — her muttering 'this can't be right' to nobody. I cut all of it. The plot beats were obvious. I rewrote them. What I kept was the Phoenix heat, the storage unit, the box, and those fogging glasses. The model gave me a setting and an object; I had to bring the story to it. That division of labor is the whole trick, and once I understood it I stopped expecting the prompt to write my book and started using it for the one thing it's genuinely great at: finding the physical world I couldn't see from inside my own head.
How I run it now, every single time I'm stuck
I no longer treat the output as a draft I keep. I treat it as a door I walk through. Here's the actual loop:
- Run it three to five times with DIFFERENT scene moments — 'the discovery', 'the lie', 'the morning after' — and read all of them before deciding which book I'm actually writing.
- Keep the setting hint vague on purpose. 'Somewhere hot and ugly' gets me Phoenix. 'A city' gets me nothing.
- Rewrite at least 40% of whatever I keep. The model's prose is a scaffold; the voice has to become mine or readers smell it.
- Steal the detail, ditch the plot. The plot it invents is usually fine and forgettable. The objects it notices are gold.
Six months after that first Phoenix scene I have 24,000 words of a novel that used to be a note about a spice rack's worth of avoidance. It is not a great novel yet. It is, for the first time, a novel that exists. If you have an idea you've been circling for years, grab the Turn a One-Sentence Logline Into a Cold-Open Scene prompt on Prompt Dock, give it your messiest logline and a deliberately vague setting hint, and let it pick a door for you. When the opening hands you forward momentum but the dialogue underneath sounds like robots, I run it straight into my Punch Up Flat Dialogue to Sound Like Real People prompt and the scene stops creaking.