Chat, Notes, Agent: The Three Stages of Using AI to Write

Three stages, and most people start at the wrong one

People use AI in three stages, and they are not really about tools. They are about what you are asking the machine to hold.

Stage one is a chat box: one question, one answer, no setup. Stage two is accumulated material: notes, a story bible, the things you have decided, kept somewhere the model can read. Stage three is a run: a task long enough that you want the machine to carry it to the end without you watching.

The stages are sequential for a reason, and the middle one is the one people skip. That skip is the most common cause of the complaint that AI drafts all sound the same — because a model with nothing but your genre label has nothing to work with except the average of everything it has read.

Stage one: a chat box and one real question

The simplest way to start using AI is to open a chat model and type. No install, no API key, no agent framework. Every hour spent configuring a local stack before you have asked a single useful question is an hour spent on the wrong problem.

The first question should be about something you actually care about, not a demo. For a writer that usually means one of these:

• "Here is my opening paragraph. What does it make you expect will happen next, and were you right to expect that?"
• "I have a character who is generous in dialogue and selfish in action. Is that a contradiction or a pattern?"
• "Explain what a scene and a sequel are, and then tell me which one my chapter two is missing."

That last shape is the most useful habit in stage one: ask for an explanation from a specific angle. "Explain this as a teacher" and "explain this as an editor who has read too many first drafts" produce very different answers, and the second one is usually the one you needed.

Stage one is free or nearly free, and it is genuinely enough for a lot of writers. If you are drafting short pieces, it may be all you ever need. The limit shows up when the work gets long: the model forgets what you told it three sessions ago, and you start repeating yourself.

Stage two: notes that answer instead of archive

This is the stage that separates writers who get something out of AI from writers who collect tools.

The digital hamster problem

Most of us keep notes. Principles, lessons, snippets, a document called ideas.md that has not been opened in eight months. It feels like progress. It is mostly hoarding — you are a hamster filling a cheek pouch with material you will never retrieve, because retrieval is the hard part.

A folder of notes cannot answer a question. Search returns files, and a file is not an answer. If your notes are scattered across eleven documents and three apps, the honest situation is that you have written them and you do not have them.

This is the one thing a language model does better than a filing system: it will read a pile of fragments and answer a question about them. Not because it is cleverer than you, but because it does not mind reading all eleven documents every time.

What a queryable bible looks like

For fiction, stage two has a concrete form: a story bible the model can be handed in full.

Not a wiki. A wiki is a set of pages you navigate — useful for a human, useless for a question like "what does Mara know about her brother at the end of chapter twelve, and what does she merely suspect?" A pile of pages cannot answer that. A single document plus a model can.

Keep it small enough to paste. Four sections, updated as you go:

• Facts. Names, dates, physical details, the rules of the world. Anything you would be annoyed to see contradicted.
• What each character knows, and when. The most useful section and the one nobody writes. Half of all plot holes are knowledge errors.
• Open questions. Threads you planted and have not paid off. This list is your outline for the middle of the book.
• Voice notes. Three sentences you wrote that sound like the book, kept as samples. Paste them when you need the model to match register.

Then every prompt starts with the bible instead of a summary. The difference is immediate and slightly unnerving: the model stops inventing a generic protagonist and starts using yours.

There is a pleasant side effect here. Once your own material is in the loop, you stop asking the model to be original. It does not need to know the world; it needs to route around the world you built. You are not archiving knowledge anymore, you are issuing instructions — and the instructions are only as good as the material behind them.

Stage three: something that finishes the run

An agent is a model with a task list and the ability to keep going: read the outline, draft chapter one, check it against the bible, draft chapter two, export the file. You set it running and come back.

This is where the leverage is for long projects. It is also where people waste the most money, because an agent pointed at a vague goal will produce a large amount of confident nothing, and it will do it for hours.

Three rules keep stage three useful:

Automate the part you have already done by hand once. If you have not manually drafted three chapters with the bible, you do not know what instructions the agent needs, and you will write them badly.

Make the output checkable. Chapter files, a word count, a contradiction list. Anything you can scan in two minutes. Long autonomous runs without checkpoints produce work you have to read in full, which cancels the time saved.

Keep a human gate at the quality boundary. The agent can produce ten drafts. It cannot tell you which one is worth finishing. That judgment is the one thing that stays with you, and it is also the only thing that separates two writers using identical tools.

The stage everyone skips

Almost everyone bounces between stage one and stage three. They chat for a while, get generic output, then go looking for an agent that will write the whole book. Somewhere in between sits the boring middle: writing down what you have decided, in one place, in a form something else can read.

That work is not fun. It does not feel like writing. It is the difference between a model that sounds like a model and a model that sounds like your book.

Where this breaks

Two boundaries worth stating before you rely on any of it.

Precise domains need primary sources. Medicine, law, contracts, tax, anything where being wrong has a cost that lands on a person. Models hallucinate, and a confident wrong answer about a dosage or a clause is not a drafting inconvenience. Use the model to understand the shape of the question, then verify against the actual document.

The model does not know your taste. It knows the average of what it has read, which is a fine starting point and a bad destination. The section that makes your book yours is the one you would have to argue for. That argument is covered in the three patterns that give AI prose away, and no amount of workflow avoids it.

FAQ

Do I need to install anything to start?

No. Stage one is a browser tab. If you are writing a fixed form and want the requirements built into the interface, the generators are a faster start than a blank chat window, because they encode the constraints you would otherwise have to explain.

How big should the story bible be?

Under two thousand words is a good target, and every section should earn its place. A bible that is too long gets skimmed by you and truncated by the model. Facts and knowledge states first; prose samples second; nothing that only matters to you.

Is stage three worth it for a single short story?

Rarely. Agents pay off on volume and repetition — forty chapters, fifty product descriptions, a series. For one short piece, the setup costs more than the draft. Start at short-form stories and see whether the workflow is even the bottleneck.