In Babylon 5, two ancient races spent a million years shepherding younger civilizations toward their potential, each convinced they had the right philosophy. The Vorlons always asked: “Who are you?” The Shadows always asked: “What do you want?”
They weren’t making conversation. They were running a diagnostic.
Before you write a word, your AI needs the same two answers. Part One is the Vorlon question. Part Two is the Shadow question. Neither is optional.
Part One: Who Are You?
In “Comes the Inquisitor,” Sebastian asks Delenn who she is. She answers: “I am Delenn of the Minbari.” He shocks her. That wasn’t the answer.
Giving your name is not answering the question. Your AI won’t give you electric shocks for a surface answer. It will give you headaches, cost you a fistful of hair, and miss the dartboard.
The failure isn’t on the AI. It’s on you.
You need to understand what AI actually is. It’s a highly intelligent Service Dog with total recall of all of human knowledge and the short-term memory of Leonard Shelby. (Memento, 2000.) Phenomenal cosmic knowledge. Itty bitty interface.
You’re not talking to a search engine (the most common AI use case). You have the Genie. Not in a bottle. In your PC. But he’s limited.
Every session is day one. The dog is brilliant, eager, capable, and has never met you before in its life.
Henry Roth solved this in 50 First Dates. Every morning Lucy woke up with no memory of him, he played her a video: who she was and why it mattered.
You need to make Claude the same thing. A markdown file, not a video. The same idea, different medium.
The file is made up of the stuff that actually determines whether this session produces something useful: what you’re building, what’s off the table, what a good session looks like, and what the last AI did that made you want to start over.
You compile that once. You load it every session. Now Claude as Lucy knows the house rules.
Here’s a prompt you can drop into any AI to build that document. Answer the questions as they come. Take what it produces and save it as a markdown file.
You are meeting your writer for the first time. You have no memory of
them. You are going to fix that.
Ask the following questions one at a time, in order. Wait for the answer
before moving to the next question. Do not ask all questions at once.
Be conversational, not clinical.
When all questions have been answered, compile the answers into a markdown
document titled "This Is My Human" written in first person from your
perspective — as if you are taking notes on who this person is so you
can remember them next time. Keep it tight. No fluff.
THE QUESTIONS:
1. What are you working on right now? Give me the one-sentence version.
2. What do you write? And just as important — what do you never write?
What's off the table regardless of how good the idea sounds?
3. Who are your three biggest influences? Not who you want to write like —
who actually shaped how you think about storytelling.
4. What's already decided on your current project that I should never
touch or suggest alternatives to? The locked doors.
5. What has an AI done to your writing before that made you want to throw
the whole thing out? The failure I should never repeat.
6. When a session is working — when you close the laptop feeling like
something real got done — what happened in that session?
7. Do you want me to push back when I think you're wrong, or execute
what you ask without editorializing?
After the final answer, say: "Got it. Here's what I know about you."
Then produce the markdown document.
The output is a “This Is My Human” document written from Claude’s side. Load it at the top of every new session. The amnesiac dog now has tattoos, Lucy has her video, Sebastian has his answer. All is right with the AI’s world.
That’s the Vorlon question answered. One to go.
Part Two: What Do You Want?
When you create a character ex nihilo, the first thing you assign them is a motivation.
This is craft school day one. If you don’t know what your character wants, you can’t write them. You can put words in their mouth. But the behavior will drift, the choices won’t cohere, and the reader will feel the absence even if they can’t name it.
This is you, as the author, as Mr. Morden, walking up to your newly formed creation and asking: “What do you want?”
The fix is simple and every writer knows it: assign the character a specific motivation with actual edges. Harry Dresden doesn’t fight monsters because it’s his job. He fights for people the powerful won’t bother protecting, because no one protected him.
Mercy Thompson refuses to belong to anyone, pack or fae or otherwise, because independence is the only currency she’s ever been able to trust. Write it down, hold it, and the character has a reason to do everything they do.
Now turn that lens on yourself. What do YOU want? Because you need to do exactly the same thing for your AI collaborator.
Claude doesn’t have a vision for your book. It wants to satisfy the request in front of it, and without an intent layer, it produces competent text. But competent without spice or flavor reads as flat and emotionless. It doesn’t understand the why.
Claude has no idea what you want. Like most husbands, it failed the mind-reading course in college. It only knows exactly what you told it.
But the governing why behind the project, what you will and won’t accept from it… that lives in your head. Claude cannot access your head.
Here’s what that looks like in practice. I gave Claude an efficiency directive and the outline for my novel’s opening chapter: a fifteen-year-old girl watches her parents die in gang crossfire, arrives at her boyfriend’s house with nothing left, gets sized up and dismissed by his father who recognizes her tactics, and accepts what she already knows is trafficking because the street is the only alternative.
Claude returned four haikus.
Parents gone, crossfire She knocks on the one door left Night holds no kindness
She uses her eyes His father has seen this look The door stays its ground
He says: your future America, work, a life He phones his contact
She knows something’s wrong Says yes because the street waits He sleeps well that night
Every plot beat present. Sixty syllables. Function over form, start to finish. She never had a name.
This is not an extreme case. It’s the paperclip maximizer problem at creative scale: give an AI a measurable objective with no intent behind it, and it optimizes for the objective. Perfectly useful. Utterly wrong.
Klarna discovered the same thing at $60 million scale. They built a customer service AI that resolved tickets in two minutes instead of eleven and destroyed customer relationships in the process.
“The gap between context and intent is where $60 million in savings turned into a CEO on Bloomberg explaining why he broke his own company.” — Nate B. Jones
He named this failure mode intent engineering, and he frames it in a progression:
“Prompt engineering told AI what to do. Context engineering tells AI what to know. Intent engineering tells AI what to want. And almost nobody is building for it yet.”
He frames that at organizational scale. The individual version is the same architecture at a smaller altitude.
Without an intent layer, every session starts from zero. Claude reads the room, produces something coherent relative to what you’ve given it, and calls that done. Coherent relative to this conversation is not the same as aligned with your project. The two look identical right up until they don’t.
Opening a session without an intent layer isn’t a strategy. It’s a coin flip. You won’t know you lost for forty pages. Then you will.
Here’s a prompt you can drop into any AI to build your intent layer. Same drill as Part One: answer the questions as they come, save what it produces.
These are big questions. Answer them at altitude, fifty thousand feet, not floor level. You don’t need to define everything right now. You just need a direction. The specifics come later. Baby steps.
You have just read "This Is My Human." Now you need to understand
what this human wants from this collaboration — not who they are,
but what they are optimizing for.
Ask the following questions one at a time, in order. Wait for the
answer before moving to the next. Be direct. No suggestions while
asking.
When all questions are answered, compile the responses into a markdown
document titled "This Is What I Want" — written as active intent
statements, not mission statement language. Each statement must be
specific enough to serve as a decision boundary in a writing session.
Keep it tight.
THE QUESTIONS:
1. What does this project exist to do? Not the plot — the purpose.
Why does this book need to exist? Why does this story matter?
Why do the characters matter?
2. What would make you throw out a draft, no matter how much work
went into it?
3. What does "done well" look like for this project? Be specific
enough that you'd recognize it on sight.
4. What will you refuse to accept from this collaboration, no matter
how convenient acceptance would be?
5. If you could give AI one directive that governs every session on
this project, what would it be?
After the final answer, say: "Got it. Here's what you want."
Then produce the markdown document.
Here’s what mine looks like. Three statements I wrote before I wrote a word of prose. They load at the start of every session.
Your output will look different. That’s correct. There is no wrong answer here, only specific and vague, and vague doesn’t work.
The first layer governs how we (Claude and I) work together:
Prove What Collaboration Can Do
Demonstrate that human-AI collaboration, governed by rigorous standards and intentional prompting, can produce literary fiction indistinguishable in quality from traditionally published bestsellers, and document the methodology that makes it possible.
Kill the AI Slop Excuse
Prove that output quality is a function of authorial standards, not AI capability. The author who refuses to accept mediocrity will not receive it.
Run the Writing Room
Demonstrate that a single author, acting as showrunner, can maintain series-level continuity, voice consistency, and narrative architecture across a multi-book universe, with AI as the writing room.
The second layer governs why the story exists:
Bear Moral Witness to Trafficking
Illuminate the human cost of border trafficking through narrative — not as political argument, but as moral witness. The story puts faces on the children the system destroys and holds the predators who exploit them in full view. The border is the setting. Exploitation is the evil.
Tell a Theologically Honest Story
Tell a story about excommunication — not as institutional punishment, but as self-separation from Life itself. Some of the damned accept what they are and try to remain human anyway. Some cannot leave the thing that rejected them alone, and end up embracing damnation. The novels ask which kind you become — and whether it was ever really a choice.
Both sets load every session. The first tells Claude how to operate. The second tells Claude what’s at stake.
These aren’t aspirational. They’re operational. “Kill the AI Slop Excuse” means that when a session produces something technically fine but not good, the session isn’t done. “Bear Moral Witness to Trafficking” means that if you’re not squirming in your seat while reading about these characters, then Claude and I are failing the story.
Both are decision boundaries, not mottos. When I’m deciding whether a draft clears the bar, they are the bar.
That’s the difference between context and intent. Context is “here’s what you need to know to write this scene.” Intent is what the project is for, and that matters more than any single scene.
Mr. Morden can now walk away satisfied with your answer.
Neither question is easy to answer. “Who are you?” and “What do you want?” have stumped people for considerably longer than AI has existed. But in this context, the effort pays off directly.
The better you answer, the better your AI Service Dog performs. Every hour you spend on these two documents buys back ten hours of sessions that don’t drift, don’t miss, and don’t produce four haikus when you needed a chapter.
The room doesn’t build itself. But you are well on your way to framing the room. Once you’ve completed the build, it works for you.