The Writer's Workflow: Where AI Earns Its Seat at Each Stage
The stack above is organised by tool. A writing project moves through stages, and the writers getting the most from AI have stopped asking "what can it write for me" and started asking "what does this stage need". Here is the map.
Ideation and research. The widest funnel of the process and AI's most frictionless contribution: twenty title angles, five structural approaches to the same argument, the research summarised with citations (Perplexity for anything that must be sourced), the premise stress-tested before a word of draft exists. The selection remains yours, and the selection was always the creative act at this stage anyway.
Drafting. The stage where AI use varies most by discipline, honestly. Fiction writers in our testing got the most from collaborative drafting (Sudowrite's Describe and Rewrite on prose that exists) rather than wholesale generation. Non-fiction writers got the most from structure-first drafting: AI expands an argued outline into rough prose the writer then rewrites into their voice. Copywriters got the most from variant volume: ten headline options beat one, every time. The constant across all three: the blank page problem dies, and the writer's energy moves from producing words to improving them, which is where craft lives.
Revision and editing. The stage where the tools have quietly become extraordinary. The structural pass (Claude holding the full manuscript: is chapter 8 consistent with chapter 2, where does the argument sag, where does the voice drift), then the craft pass (ProWritingAid's pacing, dialogue, and repetition reports), then the line pass (Grammarly, Hemingway). Three editorial layers that used to require three different humans, available the same evening the draft finishes. The judgement on which suggestions to take remains the writer's, and that judgement is exactly the editing skill it always was.
Publication and promotion. The afterlife work writers universally resent: the meta description, the social copy, the newsletter blurb, the query letter variants, the back-cover copy attempts. AI handles this layer almost completely, in your trained voice, which means the hours after "done" stop eating the hours before the next "begun".
The pattern across all four stages: AI compresses every part of the process except the part that makes the work yours.
Prompting as Craft: Directing the AI Like an Editor Directs a Writer
The single biggest variable in AI output quality is not the tool. It is the brief. Writers, of all people, should recognise the dynamic: a vague brief produces a vague draft from any writer, human or machine, and a precise brief produces something usable. Prompting is the brief-writing skill, applied to a tireless and slightly literal collaborator.
The anatomy of a working prompt. Five elements, every time: the audience (who is this for, specifically), the tone (with a reference point: "like a business advisor over coffee", "like late Didion", "like my attached samples"), the format and constraints (length, structure, what to include, what to never do), the key content (the points that must land), and examples where stakes are high. Compare "write a blog post about AI" with "draft a conversational post for small business owners on AI customer service automation, focused on cost savings, friendly advisor tone, ending with a demo call-to-action": the second produces a draft needing minutes of editing, the first produces the internet's average.
Three prompt patterns that earned permanent places in our testing:
The sparring prompt: "Here is my premise/argument/plot. Argue against it as my smartest sceptical reader. Do not be polite." The output is a list of weaknesses you fix before any reader finds them, and it is the cheapest developmental edit in existence.
The alternative-takes prompt: "Here is my paragraph. Give me five rewrites: one tighter, one more vivid, one in a different emotional register, one restructured, one that breaks the rules. Label what each is doing." You rarely take any wholesale; you steal a verb from one and a rhythm from another, which is exactly how working with a good editor feels.
The constraint prompt: "Rewrite this scene/section keeping every plot beat but cutting 40 percent." The forced compression surfaces what was load-bearing and what was habit, and the diff between versions teaches more about your own padding than any style guide.
The iteration rule. First outputs are first drafts of the AI's understanding of your brief. The craft is in the follow-up: "warmer", "less hedging", "the second paragraph is doing nothing, fix or kill it". Writers who treat prompting as a one-shot transaction get generic results and blame the tool. Writers who direct iteratively, the way an editor directs revisions, get output worth the subscription.
The Writers' Room of One: AI as Critic, Not Ghostwriter
The verdict named the biggest mistake (generate and paste) and the biggest opportunity (the thinking partner). This section is the opportunity, expanded, because it is where our testing found the genuinely transformative use cases, and none of them involve the AI writing your prose.
The critic on call. The highest-value prompt in a writer's repertoire is some version of "what is wrong with this": where does the pacing die, which character sounds like the others, where is the argument assuming what it should be proving, what would a hostile reviewer say. AI criticism is not as good as a brilliant human editor's. It is available at 11pm on a Tuesday, it never tires, and it has no incentive to be kind, which makes it dramatically better than the alternative most writers actually have, which is nobody.
The uncomfortable-possibility engine. Stuck stories and sagging arguments usually suffer from the same disease: the writer is protecting something (a character, a structure, a cherished paragraph). Asking the AI "what would happen if the thing I am protecting were the problem" (kill this character, invert this argument, cut this opening chapter) surfaces options the writer's attachment was hiding. You decline most of them. The act of seeing them declined consciously, rather than never considered, is the value.
The question-asker. Before drafting, the prompt "interview me about this book/essay/story until you understand it well enough to summarise it back" does something subtle: your answers to the AI's questions become your clearest articulation of the project so far, and the gaps you cannot answer are your outline of what is unresolved. The AI's summary at the end is a bonus; the forced articulation was the work.
The through-line, and the reason this section sits beside rather than above the drafting tools: every use here strengthens the writer's own thinking and leaves the prose-making in the writer's hands. The writers' room model scales the part of collaboration that makes work better (challenge, options, questions) without outsourcing the part that makes it yours.
Use Case Scenarios
If you are a novelist working on a literary fiction project, the right stack is Sudowrite Professional at $29 per month for drafting and ideation, Claude Pro at $20 per month for structural and dialogue work, ProWritingAid annual at $120 per year for editing, and Scrivener at $59 one-time for organisation. Total annual cost: around $700 for the working novelist stack.
If you are a self-publishing genre fiction author releasing series fiction, prioritise speed over polish. NovelAI or Inkfluence for faster drafting, plus a lighter editing pass with Grammarly and Hemingway. The stack is cheaper ($200-400 per year) but the workflow is built for volume.
If you are a non-fiction author working on a book-length project, Claude Pro at $20 per month is the foundational subscription. Add Perplexity Pro at $20 per month for research with citations, ProWritingAid for editing, and Hemingway free for clarity checks. Total around $480 per year.
If you are a working copywriter at an agency, the right stack is Jasper Pro at $69 per seat per month, Claude Pro at $20 per month for nuanced individual work, and Grammarly Premium. Total around $100 per month: high for a writing stack but appropriate for client volume and brand governance requirements.
If you are a freelance copywriter or independent marketing writer, skip Jasper and use Claude Pro plus ChatGPT Plus plus Grammarly. Total around $50 per month, covering most copy work without enterprise pricing.
If you are a journalist, Perplexity Pro is the priority subscription. Add Otter.ai for interview transcription, Claude Pro for synthesis and writing, and Google Journalist Studio (free) for archive research. Total around $50-70 per month.
If you are a newsletter writer building a paid audience, Claude Pro plus your platform (Substack free, Beehiiv has paid tiers) plus Grammarly free is genuinely enough for most newsletter operations. Total under $30 per month plus platform fees.
If you are a screenwriter, Sudowrite or WriterDuet AI plus Claude for dialogue work is the standard stack. Total around $50-80 per month depending on tier choices.
If you are a ghostwriter managing multiple clients, Claude Pro or ChatGPT Plus with Projects/Custom GPTs is the highest-leverage tool. The voice training capability across multiple client contexts is unmatched. Add specialised tools per client based on what each project requires.
If you are just starting and want to test AI writing tools, the free stack covers a meaningful percentage of writer needs: Claude free + Hemingway + Grammarly free + Google Docs. Spend money on paid tools only when you understand which specific bottlenecks they would solve in your actual practice.