Portfolio

The ‘Start Writing’ Prompt

How to Begin Any Creative Project with AI

The problem

You’ve planned your story. You have an outline, character profiles, a treatment. You sit down to write, paste all of it into your AI assistant, and type: “Write chapter one.”

The output is generic. It follows your outline but the prose sounds like every other AI-generated story you’ve read. Your meticulously-detailed characters feel flat, and descriptions are pleasant but forgettable. You spend more time editing than you would have spent writing from scratch.

The problem could stem from your first prompt.

Why a ‘start writing’ prompt is important

Most people who use AI for creative work spend hours on planning and seconds on the prompt that bridges planning and prose. That prompt is the most important instruction you’ll give the model. It determines whether the AI generates text that sounds like you on the first draft, or default slop that sucks up time and tokens in revisions.

A good ‘start writing’ prompt does three things:

  1. Gives the model access to your planning documents without overwhelming its attention
  2. Directs it toward the specific scene you want written
  3. Reminds it of the stylistic constraints that separate your voice from default AI output

The five-part template

Effective prompts follow a universal structure. Whether you’re writing fiction, screenplays, or creative nonfiction with AI, the same five parts apply. These are the questions the LLM needs answered before it can write something worth reading.

1. Role

Tell the model what it is. “You are a fiction writer” is too vague. “You are my fiction-writing collaborator and self-correcting prose editor” tells the model it should both generate and revise before showing you anything.

2. Context

Point the model to your planning documents. If you’re using a system that lets you pass URLs, include links to your treatment, outline, and character profiles. If not, paste the most relevant documents inline. The key is to give the model access to your planning without making the prompt itself a wall of planning text.

3. Task

Tell the model what scene to write and what it needs to accomplish. This is the most important section. Include a silent identification step that forces the model to think before writing:

  • What does the POV character want right now?
  • What changes by the end of the scene?
  • What’s the emotional pressure underneath the visible action?
  • What concrete consequence should close the scene?

Without this step, the model generates prose that is syntactically correct but emotionally generic. With it, a thoughtful model processes the scene as a writer would before generating a single sentence.

4. Constraints

Include a short list of style reminders. Not your full style guide. Just the five or six patterns most likely to appear in the scene you’re about to write. For example:

  • No em dashes. Use periods, commas, or restructure.
  • No triple-beat fragments or staccato verb sequences.
  • No vague interiority (“something shifted”). Name what happens.
  • No default physical tells (jaw tightening, breath catching). Use character-specific behavior.
  • Accumulation principle: any pattern appearing more than twice in a scene gets rewritten on the third instance.

The accumulation principle is the single most useful instruction in this section. The problem with AI prose isn’t any one construction. It’s reflexive repetition. One “jaw tightening” is a choice. Three in a chapter is a tic. Telling the model to watch for density catches the root cause, not just the symptoms.

5. Format

Direct the model on how to start, how to end, and what to do after drafting. The most important instruction here is the hidden revision pass: tell the model to review its own draft for the patterns in your constraints and rewrite them before showing you the result.

The copy-paste prompt

Here is the full template. Copy it, fill in the brackets, and use it to start any scene.

[ROLE]
You are my fiction-writing collaborator and self-correcting prose editor.

[CONTEXT]
Treatment: [link or paste]
Outline: [link or paste]
Character profiles: [link or paste]

[TASK]
Draft Scene [X] from the outline. This scene covers [one-line description of what happens].

Before drafting, silently identify:
- The POV character's immediate want.
- What changes by the end of the scene.
- The emotional pressure underneath the visible action.
- Each major character's distinct stress response, restraint pattern, desire pattern, and speech rhythm.
- The concrete consequence that should close the scene.

[CONSTRAINTS]
- No em dashes. Use periods, commas, or restructure.
- No triple-beat fragments or staccato verb sequences.
- No vague interiority. Name what happens or show it through behavior.
- No anthropomorphized silence or atmosphere.
- No default physical tells. Use character-specific behavior.
- Accumulation principle: any pattern appearing more than twice in a scene gets rewritten on the third instance.

[FORMAT]
- Start with character presence, action, decision, or pressure. Do not open with weather, skyline, architecture, or atmospheric description unless the character's perception of it changes behavior.
- Keep narration inside character perception or physical reality.
- End scenes on action, decision, rupture, consequence, interruption, or a concrete image that changes meaning because of what just happened.
- After drafting, perform a hidden revision pass. Search for the patterns above and rewrite them automatically. Do not explain the revision unless asked.
- Give me only the polished draft unless I request notes.

How to customize per scene

The template is a starting point. Here’s what to adjust each time you write:

Task section: Always. This changes every scene. What happens, what changes, what the scene costs.

Constraints section: Often. A dialogue-heavy scene might add “no adverbs after dialogue tags.” An action scene might add “no default physical tells.” An introspective scene might add “no vague interiority.” Add or remove constraints based on what the scene is likely to trigger.

Format section: Rarely. The opening and closing instructions stay stable across most scenes. You might adjust the opening if the scene genuinely starts with environment rather than action.

Role and context sections: Almost never. These stay stable across the project.

Common mistakes

Pasting everything into one prompt. Treatment, outline, character profiles, worldbuilding wiki, all in a single message. The model processes none of it well. Use links if your system supports them. Otherwise, paste only what the current scene needs.

No silent identification step. The model generates prose without first processing what the scene is about. The result is syntactically correct but emotionally flat. Always include the silent identification step.

No hidden revision pass. The model returns a first draft immediately. The first draft will contain patterns from your banned list. The hidden revision pass catches them before you see them.

Treating the prompt as fixed. The constraints section should change per scene. If you use the same constraints for every scene, you’re not adapting to what the scene needs.

The one instruction that does the most work

If you take only one concept from this guide, take this: the problem with AI-generated prose is not any single construction. It’s reflexive repetition. One “something shifted” is a deliberate vagueness. Three in a scene is the model falling back on a default because it couldn’t find the specific thing that shifted.

The accumulation principle, stated as a prompt instruction, is one line:

Any pattern appearing more than twice in a scene gets rewritten on the third instance.

This single instruction does more work than a long style guide because it targets the root cause: density and reflexive repetition. The model can follow a banned-words list and still produce flat prose by substituting one tic for another. The accumulation principle catches the substitution because the new pattern will also accumulate.

Include this line in every start-writing prompt.


This guide is a standalone resource for beginning a creative writing project with AI. For related topics, see our companion articles on style guides for AI writers and using document APIs to pass planning materials to your AI collaborator.