I’ve been poking at ChatGPT and Gemini for a few months now, experimenting with using AI to write fiction – helping with everything from brainstorming, plotting, character design and (gasp) even drafting. The goal in all this was to both explore the capabilities of AI and see how it can help my limited human attention in “visualizing” for artistic work by giving my overworked neurons something to stew on.

The short of it: AI is incredibly helpful and useful, enabling me to rapidly flesh out multiple novels in tandem, while also aiding in hammering out many detail-specific aspects to produce formidable prose.

An example of what I mean: I was writing a supernatural mystery-comedy in which the protagonist (a man) turns into a woman. Needless to say, there was a scene where the detective went shopping for a new wardrobe, and it turns out AIs are more knowledgeable about bras than Charlie Sheen, and most certainly my humble self.

Below I’ll turn my notes into a friendly, practical playbook you can steal, remix, and exploit for your own novels — and I’ll nerd out on the knobs that actually change the results (temperature, top-p, a few other useful controls).

tl;dr

Use AI to plan and seed drafts, not to produce finished voice-driven prose. Do pre-planning (characters, 5-act plot, chapter outlines) with higher creativity settings; write scenes in bite-sized, directive prompts; use lower-creativity settings for tighter, factual passages. Learn to tweak temperature and top-p — they’re simple but powerful. For authoritative docs, check Google’s Gemini/AI Studio and OpenAI’s prompt guides. 

Why use AI for fiction?

Because brainstorming ideas takes time and tremendous cognitive load. Working with an AI is like spitballing with a person – actually, it’s way better. No, really: most people suck at being creative. They’re more concerned with basic and material things like money, entertainment, and of course, sex.

In hearty contrast, AI is a brilliant brainstorming machine. It comes up with ideas that never crossed my mind and gets my blood a-pumping and my neurons a-firing. Plus, it never gets tired of my stupidity, and best of all, it never takes things personally. It is a dream come true.

Some of the things AI is useful for:

• vomiting lots of alternatives when you’re stuck (names, character beats, scene hooks),
• outlining and turning a messy idea into a structured plan,
• producing technically competent draft material to riff off of.

What AI is not (at least, reliably)

AI is not an automatic source of consistent, soulful voice. Machine prose is often technically solid but a little bland or “voiceless” — it reads like someone who knows grammar but is emotionally guarded. That’s why I edit AI-written material heavily, or more preferably, rewrite from scratch, allowing my brain the transform the AI prose into something human and palpable.

You can get great outputs on free tiers (Gemini via AI Studio, etc.), but the paid pro models and research/deep-thinking modes often add throughput or features (longer reasoning passes, retrieval, and other research tools) that speed up big-planning tasks. Those tradeoffs are worth evaluating for your workflow, especially for grinding out all the character and plot details.

My process for drafting with AI

1. Pre-planning — characters, tone, themes, and a 5-act plot: store in a Google Doc. Character planning — ask the model to produce archetypes, mannerisms, physical details, backstories. Refine until you know the characters intimately.

2. Plot & chapter plan — turn the plot into chapter-sized beats (I aim for chapters of ~3k–5k words; estimate 30 minutes reading each).

3. Scene writing (bite-sized) — hand the model short, explicit scene prompts (single scenes, single beats, or dialogue-only chunks).

4. Iterate & humanize — copy the draft into your own editor and rewrite for voice, consistency, and subtext.

This is the process that reliably accelerates my creativity while keeping control in my hands.

Prompt examples (copy/pasteable)

Character planning

Plan out the characters for a fantasy novel. I want the main character to be a dark antihero with a troubled past. List supporting characters using common archetypes: his military friends, a princess who becomes a supporting character, the royal family. For each character give: name (Latin/Roman-origin), a one-line role, 5 distinct mannerisms, 2 physical details, and a 50–100 word history.

Plot

Plan the plot for a fantasy novel in 5 acts where the antihero discovers a plot to overthrow the royal family. He fails to stop the coup but saves the princess and flees. Use common tropes where useful, and shape the arc so the hero redeems himself by the end. Make a chapter-by-chapter outline for ~80k words, with each chapter ~3k–5k words.

Scene (bite-sized)

Write a scene where the antihero overhears three conspirators planning the coup. Tone: tense, terse. Focus on sensory detail and one reveal that changes the hero's next choice. Keep it to ~700–900 words.

Side note: whole-chapter prompts work, but they tend to drift (name swaps, detail drift, pacing issues). Tiny, directive prompts reduce drift and keep the AI playing the role you intend (creative director > ghostwriter).

The knobs that matter: temperature, top-p, and why they’re useful

These two are the primary “creativity” controls you’ll see in AI UIs.

Temperature (the randomness knob)

What it is: a scalar applied to the model’s token-selection softmax that controls randomness. Lower → more deterministic; higher → more random. Many APIs default to 1.0, and some UIs allow 0.0–2.0 (Gemini typically uses that range).  Practical effect: at low temperatures (0–0.3) the model repeats high-probability continuations — great for factual, consistent text. At mid temperatures (0.6–1.0) you get more color and variety; at high temps (>1.0) you can get surprising, novel but sometimes incoherent text. Nuance: temperature is not a magic “creativity = temperature” mapping. Recent work shows the relationship is more subtle — higher temperature increases novelty a bit but can hurt cohesion. Use it as one axis of control, not the only technique for creativity. 

Top-p (nucleus sampling)

What it is: instead of sampling from the entire probability distribution, the model restricts choices to the smallest set of tokens whose cumulative probability ≥ top-p. Lower top-p = fewer candidate tokens considered.  Practical effect: top-p controls the breadth of the sample set. You can use top-p or temperature (or both), but many docs advise changing one at a time to isolate effects. 

Common Heuristics

Brainstorming / Ideation: temperature = 0.8–1.2, top_p = 0.8–0.95 — looser, more varied results.

Structured planning / outlines: temperature = 0.3–0.6, top_p = 0.8–1.0 — keeps you on track.

Tight copy or specific facts: temperature = 0–0.2, top_p = 0.5–0.9 — minimize hallucinations. These ranges are heuristics — test and adapt to the model you’re using.

Google’s docs and Vertex/AI Studio show similar guidance and range defaults. 

Other parameters and prompt-engineering patterns worth knowing

Max tokens / length — cap how long the model can respond. For chapter seeds, let the model produce larger outputs; for scene prompts, tighten the cap.

System message / role priming — use a system instruction to fix voice/role, e.g., “You are a noir fantasy novelist who favors short sentences and interior monologue.” This helps maintain stylistic consistency across prompts. (Gemini/Chat UIs give you a system field.) 

Few-shot examples — show the model 1–3 examples of the style or structure you want; it mimics them.

Chain-of-thought / thinking modes — some paid or research modes (labelled “thinking” or “deep research” in certain UIs) run longer internal reasoning or bring external retrieval; use those for big, research-heavy planning.

Frequency / presence penalties — reduce repeated tokens or concepts (handy if the AI obsesses over a single word or weirdly repeats the same sentence pattern).

For more on practical prompting patterns, OpenAI and cloud vendors publish solid best-practice guides. 

A sample session — step-by-step (my real flow)

Session A — world & cast (creative)

Set temperature ≈ 1.0, top_p ≈ 0.9.

Prompt: “Make a list of 12 character archetypes for a dark-fantasy kingdom. For each, give a core secret that could be used as a plot twist.”

Result: 12 seeds. Save the doc, prune favorites.

Session B — plot skeleton (semi-structured)

Set temperature ≈ 0.6, top_p ≈ 0.9.

Prompt: “Turn the top 6 seeds into a 5-act plot. Make act break points obvious and include a ‘reversal’ in Act 3.”

Result: a fleshed skeleton I can argue with.

Session C — chapter plan (precise)

Set temperature ≈ 0.3–0.5, top_p ≈ 0.8.

Prompt: “Expand act 2 into 10 chapter beats. Each beat: 3–6 sentences, with stakes and a small sensory anchor.”

Result: a chapter outline I can translate to scene prompts. Session D — scene writing (directive) Lower temp for consistent detail (0.2–0.5).

Prompt single scene: “Write a 800-word scene where [specific characters] do [specific action], preserving the name ‘Cassian’ and the wound on his left shoulder.” Edit, paste into my doc, humanize voice.

This structure keeps the model from drifting and keeps me in control of details.

Common pitfalls and how to fix them

Name/Detail drift: include the character’s exact name or a short character table in the prompt. Use the system message to remind the model of constants.

Tone/music mismatch: give a short exemplar sentence (“Write in the voice of X: ‘…example…’”) or a 1-paragraph exemplar.

Overlong chapters that meander: break tasks into scenes or beats. Use max_tokens to enforce size.

Hallucinated facts or impossible “research”: for real facts, either keep temperature near zero or pair the model with a retrieval system (some pro/research modes provide that).

When to go paid (and what you get)

Paid tiers often add:

• longer context windows (important for big bibles and long works),
• specialized “thinking” or “research” modes that run longer reasoning or web retrieval,
• higher throughput and fewer rate limits.

If you’re doing massive worldbuilding, running many iterative drafts, or want retrieval-augmented consistency, evaluate pro features. Otherwise, free tools (Gemini via AI Studio, ChatGPT free) are perfectly useful for a lot of the work. (I use both: Gemini for a certain technical, rhetorical flair; ChatGPT for lighter, comedic pieces — subjective tastes only.)

A short manifesto for getting better results

Use AI to amplify your best ideas, not to replace them. Plan in the AI; write (voice) yourself. The AI is a compositional power tool, not the author. Learn the knobs (temperature/top-p, max tokens, role priming) and keep a small cheat sheet of settings for different tasks. Iterate with small prompts and tiny, specific outputs rather than asking for a 10k chapter in one shot.