Using ChatGPT and Claude for Creative Writing That Does Not Suck

Why Most AI Writing Is Bad

LLMs have one fatal tendency: they regress to the mean. Ask for a blog post, you get the average of every blog post they trained on. Ask for a story, you get a story that sounds like everyone else’s story.

This is not a bug you can prompt your way out of. It is what the tool is. Getting good writing from an LLM requires you to supply the non-average parts yourself.

The Two Usable Modes

There are exactly two ways to get value from LLMs for creative writing:

Mode 1 — Thinking Partner. You are working on a piece. You hit a wall. You paste what you have and ask specific questions. What is the argument? What is weak? What would sharpen this? You take the useful parts of the answer and keep writing yourself.

Mode 2 — Mechanical Assistant. You know exactly what you want. You need a draft of a section, a rewrite of a passage in a different tone, a summary of a long source. You direct the tool with tight constraints and edit the output.

Mode 1 improves your writing. Mode 2 saves you time. Any other use case — “write me a viral Twitter thread about X” — will produce generic slop.

Claude vs ChatGPT for Writing

After two years of both, the differences are real but small.

Claude (especially Opus models): Better long-form coherence, more nuanced tone control, better at holding a specific voice across multiple prompts, stronger at nuanced editorial feedback, resists becoming sycophantic longer.

ChatGPT (4.1 and o-series): Faster iteration, better at structured output and templates, superior inline image generation, stronger at technical writing and documentation.

For serious creative work, most professional writers prefer Claude. For speed and versatility, ChatGPT. Using both is fine. Do not get religious about it.

Prompts That Actually Work

For editorial feedback

Here is a draft essay. Read it, then tell me: what is the single strongest argument? What is the weakest? Where does the voice drop? Quote specific lines. Do not rewrite.

This works because you have bounded the output. Without the “do not rewrite” clause, the LLM will helpfully rewrite your piece into something blander.

For generating specific sections

I am writing a piece about [topic]. I have outlined the structure below. Draft section 3 only, in this tone: [paste 2 paragraphs of your own writing as reference]. Max 300 words. No headers, no bullet points, no “in conclusion” endings.

Notice the explicit bans. LLMs default to bulleted, header-heavy, conclusion-loving output. You have to cut it off.

For punching holes in your thinking

Here is my argument: [paste]. Steelman the strongest possible counterargument. Who disagrees and why?

This is where LLMs actually add value. They have seen the counterarguments, they can name them, and they have no ego about your position.

For research and synthesis

Summarize the key claims in the attached PDFs. For each claim, note who is making it and whether it is contested.

LLMs are excellent at summarization and comparison when you give them actual sources. Do not ask them to generate facts from memory — that is where hallucinations happen.

The Voice Problem

Every LLM has a default voice: helpful, balanced, slightly corporate, ends with a summary, uses “moreover” and “furthermore,” loves three-item lists.

You cannot eliminate this voice. You can only suppress it with explicit instructions and reference samples.

The minimum voice brief:

• Paste 500-1000 words of your own best writing as reference
• Explicitly name the patterns you do not want (bulleted lists, concluding paragraphs, “it is worth noting,” em-dashes as a tic, “not just X but Y” formations)
• Ban the AI-tell phrases directly: “delve into,” “tapestry,” “testament to,” “dive deep”

This gets you from generic-LLM-voice to something closer to your voice. It never gets you all the way.

The Traps

The confidence trap. LLMs state wrong things with the same confidence as right things. Verify every fact, quote, and attribution before publishing.

The smoothing trap. Ask an LLM to rewrite your rough paragraph and you will get a smoother version. Smoother is not better. The rough version probably had more personality.

The expansion trap. Ask for 1500 words, get 1500 words. Most of them padding. The good version of the piece was 400 words.

The sycophancy trap. LLMs want to agree with you. Ask “is this argument strong” and you will hear yes more often than you should. Ask “what is wrong with this argument” instead.

What Good Writers Actually Do

The writers I know who use LLMs productively share the same workflow:

1. They write a rough draft fully themselves
2. They paste the draft with specific questions (what is working, what is unclear, what is missing)
3. They revise based on the feedback they agree with
4. They sometimes use the tool for mechanical passes — cutting length, fixing flow, generating headline variants
5. They never let the tool write a draft from scratch

This workflow produces writing that sounds like them, only better. The other direction — start with LLM output, edit down — produces writing that sounds like generic LLM output, only cleaner.

The Future

Writing with AI will become unremarkable the same way writing with spellcheck is unremarkable. What will matter is the underlying craft, editorial judgment, and voice. The tools amplify what you bring. They do not substitute for it.

If you have nothing to say, AI tools will help you say nothing faster.