Eighteen months ago I was charging $0.10 a word and pitching twenty cold emails a day to clear $2,200 a month. Last April I billed $8,140 from four retainer clients and worked roughly twenty-two hours a week. The difference wasn't a course, a coach, or a niche pivot. It was a tighter AI stack and the discipline to use it like a junior staff, not a slot machine.
This is the actual setup. Tools I pay for, prompts I reuse, and the parts I refuse to automate because they're the reason clients keep paying.
The Stack, And What Each Tool Actually Does
I run four paid subscriptions. Total monthly cost: about $87. Everything else is free or already inside Google Workspace.
| Tool | Monthly Cost | What I Use It For |
|---|---|---|
| Claude Pro (Anthropic) | $20 | Long-form drafts, brand voice matching, editing passes |
| ChatGPT Plus | $20 | Research synthesis, outlines, image briefs via GPT-4o |
| Descript | $24 | Turning client interview calls into raw source material |
| Grammarly Premium | $12 | Final pass for tone consistency and clichés |
| Perplexity (free tier) | $0 | Sourced research with citations I can verify |
Claude is the workhorse. Its longer context window means I can paste a client's last six newsletters, their style guide, and three competitor pages into one conversation without it forgetting the brief by paragraph four. ChatGPT is faster for short tasks and better when I need image direction for a designer. I keep both because they fail in different places.
Why I Stopped Using Jasper, Copy.ai, And The "All-In-One" Crowd
I tried Jasper for four months in 2024. The output was fine. The problem was that purpose-built copywriting tools optimize for speed of first draft, which is the cheapest part of the job. Clients don't pay $2,500 a month for first drafts. They pay for judgment: knowing when a headline is too clever, when a CTA needs a third revision, when the founder's voice has drifted.
General-purpose models like Claude and GPT-4o give me a smarter junior. Specialized tools gave me a faster intern. I needed the former.
The Workflow That Turns A $400 Project Into A $2,000 Retainer
The money isn't in writing faster. It's in offering something a client can't easily buy from a Fiverr seller using the same models. Here's the actual sequence I run for a new retainer:
- Discovery call, recorded in Descript. Forty-five minutes. I ask about customers, objections, and the last three things that didn't work.
- Transcript into Claude with a strategy prompt. I ask it to extract themes, recurring phrases the founder uses, and three angles the brand hasn't covered.
- Perplexity research pass. I verify any claim, stat, or competitor reference. I never let a model invent numbers in client work. Once burned, permanently cautious.
- Drafting in Claude with the voice file loaded. First draft in roughly twenty minutes for a 1,200-word piece.
- Human editing pass. This is where I earn the invoice. I cut roughly 30% of what the model produced, rewrite every opener and closer by hand, and add the one specific detail from the discovery call that no AI could have known.
- Grammarly for surface polish, then send.
Total time per piece: about ninety minutes. I charge $450 per article on retainer, four articles a month per client, four clients. The math works because step five is non-negotiable. The moment I skip it, the work sounds like everyone else's.
What I Refuse To Automate
Three things stay manual. Client communication, including every Slack message and proposal. Strategy calls. And the final read-through, done out loud, before anything ships.
According to a 2025 Anthropic usage report, the most successful Claude users in professional writing categories use the tool for roughly 60-70% of a task and human-edit the rest. That matches my experience exactly. Push past 80% AI and quality cracks show. Stay under 50% and you're leaving margin on the table.
FAQ
Do clients know you use AI?
Yes. I tell them in the proposal that I use Claude and ChatGPT as drafting assistants and that final work is human-edited. Nobody has objected. Hiding it is the fast track to a reputation problem.
What about AI detectors?
I don't optimize for them. They're unreliable and most B2B clients don't run them. If a client cares, I show them the discovery call recording and the editing diffs. That ends the conversation.
Could a beginner replicate this in 2026?
The tools, yes. The judgment, no — and that's the actual asset. Spend the first six months writing without AI, then layer it in. Otherwise you can't tell when the model is wrong.