Which AI chat is best for writing?

The Wrong Question: Why "Best AI for Writing" Has No Single Answer

You've probably typed "which AI is best for writing" into a search bar more than once — and gotten back three confident answers that all contradict each other. The real dilemma isn't choosing wrong; it's asking a question that has no single answer.

The most credible hands-on comparison of the big three — ChatGPT, Claude, and Gemini — comes from UC Berkeley's information school, and its central conclusion is refreshingly honest: "The honest answer to 'which LLM should I use' is not which model is best, but which model is best for what." In their testing, ChatGPT produced the most structured, documentation-ready output, Gemini was most technically precise with a real-time research edge, and Claude was the most interpretive and analytical.

That task-based framing matches what writers actually do in practice. A survey of working marketers found AI is used across the entire writing process — not just drafting:

• Long-form drafting and outlining
• Editing, headlines, and meta information
• Interview summarization and research
• Social posts and site copy

Each of those jobs rewards a different tool. The same survey found ChatGPT dominating adoption — 90% of tool-usage responses mentioned it — while Claude showed up in sophisticated multi-step workflows, like loading tone-of-voice documents into a project and layering AI outlines under human-edited drafts.

Here's the uncomfortable part: the evidence base for "best" is thinner than the discourse suggests. No independent head-to-head benchmark of writing quality exists in the current research. The Berkeley comparison, the strongest independent test available, evaluated data science tasks — not prose. Vendor testimonials claim transformative results, but independent survey data shows the biggest complaint among the 83% who publish AI-assisted content is the hours of editing required to make output publishable.

That gap between claims and verified performance is exactly why we're building SkillzBench — an independent lab for testing and benchmarking AI agent capabilities, so "which tool works" can eventually be answered with data rather than vibes. Until then, the honest answer to "which AI chat is best for writing" is a better question: best for which writing task, and with what oversight? The rest of this article breaks it down exactly that way.

The Big Three, Matched to Writing Tasks: ChatGPT, Claude, and Gemini

Ask three writers which AI chat is best and you'll likely get three different answers — and the evidence suggests all of them are right. The most credible independent comparison, from UC Berkeley's hands-on testing, concludes that the real question isn't "which model is best" but "which model is best for what."

ChatGPT is the adoption default. In a survey of 53 marketers, 90% of tool-usage responses mentioned ChatGPT, making it the runaway leader among the "big three" identified by practitioner research. Berkeley's testing found it produces the most structured, documentation-ready output — the strongest choice when the goal is a structured, shareable deliverable. It also has the most mature community support, documentation, and third-party integrations of the three.

Claude is the power user's collaborator. Berkeley's testers described its output as reading "the way a senior data scientist might talk through findings with a collaborator" — interpretive and analytical rather than formulaic. That strength shows up in real workflows: one survey respondent described a six-step professional process built around Claude, including transcript cleanup, loading tone-of-voice and strategy docs into a Claude project, AI-generated outlines, and human-edited first drafts. If your writing depends on multi-step, document-heavy workflows, Claude fits naturally.

Gemini wins on technical precision and real-time research. Thanks to Google Search integration, it holds a real-time information advantage — useful when accuracy and currency matter more than prose polish.

On the practical details, the three are closer than you might expect:

• API pricing (per million input tokens): Gemini 3.0 Pro at $2.00, Claude Sonnet 4.6 at $3.00, and GPT-5.5 at $5.00 input / $30.00 output.
• Context windows: all three offer 1M-token context windows at the API level; GPT-5's consumer-facing window is 272,000 tokens.
• Data retention: OpenAI retains data up to 30 days, Google up to 55 days, and Anthropic just 7 days with no training use by default — a meaningful difference for confidential client work.

One caveat deserves emphasis: no independent head-to-head writing-quality benchmark exists in the available research. Berkeley's testing focused on data science tasks, not prose. And whichever tool you choose, the workflow matters more than the model — the biggest complaint among the 83% of marketers who publish AI-assisted content is the hours of editing required to make output publishable, and hallucination remains a documented risk (one respondent reported AI inventing an interview subject who didn't exist).

That's also why customization is trending: writers increasingly load tone-of-voice docs and instruction packages into their AI projects. At SkillzBench, we're building exactly that verification layer for agent skills — benchmarking which instruction packages actually lift output quality, since published research suggests roughly 26% carry at least one vulnerability. Skills, proven — join the waitlist to follow the benchmarks as they publish.

Why AI Writing Still Needs a Human Editor — The Quality Problem

Ask any AI chat to "write me a blog post" and you'll get something publishable in ninety seconds. Ask the people actually publishing that content, and a different picture emerges: 83% of marketers now publish AI-assisted writing, but their single biggest complaint is the hours of editing required to make the output usable, according to a practitioner survey of 53 marketers.

That gap between vendor promises and working reality is the quality problem nobody sells you a solution to. Market analysts identify hallucination as a top restraint on the AI writing market, noting that factual errors and inconsistent tone threaten brand credibility when AI drafts external content (Mordor Intelligence). The survey puts a face on the risk: one respondent reported AI inventing an interview partner for a customer case study — a source who simply doesn't exist.

The tension is stark. Vendor testimonials promise transformation, while the same survey captures practitioners describing AI prose as "that of an eager grad student. And oh so bland." One professional put it bluntly: "My best writing comes from not using AI at all."

The practitioners getting real value have quietly adopted what screenwriter David Cornue calls a creative firewall: "Do not write this for me. I'm the writer. You are not the writer." In that framing, AI never drafts — it supports:

• Editor: tightening your draft, flagging weak transitions, pressure-testing arguments you've already made.
• Researcher: summarizing interviews, surfacing background, and checking structure before you commit to a position.
• Wordfinder: replacing the thesaurus-and-tab-hopping loop when a sentence almost works but not quite.
• Outliner: organizing your thinking into scaffolds you then write against, in your own voice.

Notice what's missing from that list: ghostwriting. Even OpenAI's own use-case page features professionals using ChatGPT this way — as a collaborator, not an author. And the workflow evidence backs it up: the most sophisticated Claude users in the survey built multi-step processes where AI handles transcript cleanup and outlines, but humans write the first drafts.

This is also where tool choice matters less than setup. Writers increasingly customize their AI assistants with instruction packages — tone-of-voice documents, style guides, task-specific rules — to keep output on-brand and reduce editing time. But those instruction packages are only as trustworthy as their contents; published research has found roughly 26% of agent skills carry at least one vulnerability. That's the problem SkillzBench is being built to solve: verifying, security-scanning, and benchmarking agent skills so writers can adopt them with confidence rather than guesswork.

The takeaway isn't "don't use AI for writing." It's that the models are best at everything surrounding the writing — and the writing itself still belongs to you.

Prompting and Workflow: How to Actually Get Good Output

The difference between a writer who loves AI and one who spends "hours of editing" on bland output usually isn't the chatbot — it's the prompt. According to a survey of 53 marketers, the biggest complaint among the 83% who publish AI-assisted content is the editing time required to make output usable. That gap is where prompting technique does the heavy lifting.

Atlassian's four-part framework — persona, task, context, format — works in any AI chat. As Jamil Valliani, Atlassian's Head of Product for AI, puts it: "Open-ended prompts usually result in more open-ended results, which means you have less control over your output" (Atlassian's prompt guide). Tell the model who it is, what to do, what it needs to know, and what shape the answer should take — then refine iteratively rather than restarting from scratch.

The strongest practitioners treat AI as a collaborator across the whole process, not a ghostwriter. The survey found writers using AI for outlining, editing, headlines, research, and interview summarization — not just drafting. One respondent described a six-step workflow built around Claude: transcript cleanup, loading tone-of-voice and strategy documents into a Claude project, AI-generated outlines, and human-edited first drafts.

A practical version of that workflow looks like this:

• Clean and load your source material, style guide, and voice documents into a project or persistent context
• Generate an outline, then edit it yourself before any prose exists
• Draft with a firewall — screenwriter David Cornue's rule: "Do not write this for me. I'm the writer. You are not the writer" (OpenAI's use-case page)
• Use AI as editor, wordfinder, and devil's advocate on your draft
• Fact-check everything — one respondent caught AI inventing an interview partner who never existed

That last step matters because hallucination remains a top market restraint, with factual errors threatening brand credibility when AI drafts external content (market research shows a -3.8% CAGR drag from this risk alone).

If you're extending this approach with agent skills and instruction packages — the same way writers load tone-of-voice docs into Claude projects — verification matters: published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is building the benchmarking and security-scanning infrastructure to test which instruction packages actually improve output, so writers can adopt them with confidence. Skills, proven — join the waitlist to hear when results go live.

Leveling Up: Verified Agent Skills for Serious Writing Workflows

The real divide among serious writers isn't which chatbot they use — it's how much work they've put into customizing it. The most productive workflows in a survey of working marketers weren't simple prompts; they were multi-step systems, like one respondent's six-step Claude process that loaded tone-of-voice and strategy documents into a project before a single word was drafted.

That approach is spreading fast. Writers increasingly package their style guides, brand voices, and editorial rules into instruction files — agent skills — that work across tools like Claude Code, Codex, Cursor, and Gemini CLI. The market reflects the shift: the AI writing assistant space is projected to grow from USD 1.77 billion in 2025 to USD 4.88 billion by 2030, and enterprise buyers now expect brand-style-guide training and multi-agent orchestration as standard features.

But there's a problem nobody in the tool-comparison conversation talks about: skills are code-adjacent artifacts, and they're risky. Published research found that roughly 26% of skills carry at least one vulnerability — and most writers installing them from GitHub have no way to know which ones. Meanwhile, hallucination remains a top market restraint precisely because, as one survey respondent put it, AI can "invent" things that don't exist, and the biggest complaint among the 83% who publish AI-assisted content is the hours of editing required to make output usable.

This is the gap SkillzBench was built to close. As an independent testing lab for agent skills, it's building a pipeline that treats skills the way writers wish all AI output were treated: with verification before trust.

• Discover and verify — continuously crawl skills from GitHub, pin every version to a commit SHA, and run license checks, security scans, and quality judgment on each one.
• Benchmark the Lift — test every skill in a sandbox with the same model, with and without the skill, and publish the difference in percentage points. Lift is always shown separately from ratings, and money never touches scores.
• Equip agents safely — one MCP server plus per-agent loadouts, so each agent only sees the skills and tools you deliberately chose for it.
• Stay current — track upstream changes, re-scan, re-benchmark, and notify you when a skill you rely on shifts underneath you.

For writers, this changes the question from "which AI chat is best?" to something more useful: "which verified skills actually improve my chat, and which ones put my workflow at risk?" The honest answer to the first question, as one academic comparison concluded, is that no single model wins — but a verified, benchmarked skill stack can make any of them better at your specific voice.

SkillzBench is in active pre-launch development, and the waitlist is open now. If you want to know which skills genuinely help before you install them — join the waitlist. Skills, proven.