Last spring a friend of mine, a freelance copywriter, spent four hours getting an AI-generated draft ready to publish. The draft itself took about ninety seconds. The rest of the time went into the parts nobody shows in demo videos: cutting a throat-clearing opener, replacing two statistics the model invented, and untangling a paragraph that said the same thing three times in slightly different words.
That ratio of ninety seconds to four hours is the most useful thing to understand before you start. Generating is cheap. Finishing is the work. What follows is the exact sequence I use to get from a blank page to a publishable piece, with the numbers and examples pulled from real jobs.
Step 1: Write the Brief Before You Touch the Model
“Write me a blog post about email marketing” is not a brief. It’s a coin flip. The model has to guess your audience, your angle, your credibility markers, and how the piece should land. It guesses wrong most of the time, and then you blame the tool.
Spend fifteen minutes on a one-page brief instead. Mine looks like this:
- Audience: e-commerce founders doing $500k-$5m a year, already running email but not segmenting.
- The one job: convince them to split their list into three segments before their next campaign.
- Points only I can make: the 11% revenue lift we saw for a skincare client after segmentation, the list hygiene habit quietly wrecking their deliverability, the specific three-segment split I recommend.
- Tone reference: Shopify’s blog, plain and practical, no hype.
- Hard nos: no “in today’s digital landscape”, no exclamation marks, no fake urgency.
That brief does two things. It forces you to decide what the piece is actually about, and it gives the model enough friction to stop producing oatmeal.
Step 2: Generate Three Drafts, Not One
A single prompt gets you a single interpretation. Three prompts cost you nothing extra and give you options.
Vary the constraint, not just the wording
For a recent landing page I ran the same brief three ways. Draft one: “write this as a 400-word skeleton, bullet points only.” Draft two: “write it as a skeptical industry veteran who has watched segmentation projects fail.” Draft three: “write it for a founder who tried this last year and got no results.”
Draft one gave me structure. Draft two gave me the objections section, which became the strongest part of the final page. Draft three produced the single best line of the whole project, about how a segmented list nobody writes to is just a tidier version of the same problem.
Step 3: Run the Frankenstein Pass
Now you assemble. Take the skeleton from draft one, the objections from draft two, that one line from draft three. Delete everything else. In practice I keep maybe 20% of the generated text, and I shuffle the surviving pieces so they don’t sit in the order the model produced them.
This is also where you fix the rhythm. Models default to paragraphs of roughly equal length and sentences of roughly equal complexity, which is the main reason AI-generated writing feels hypnotic in a bad way. Break it up. A four-word sentence after a thirty-word sentence reads like a person. Eight medium sentences in a row reads like a machine.
If you want a fuller version of this assembly process, the six-step generative AI workflow walks through the same idea with different examples.
Step 4: Put Your Own Facts Back In
Language models are fluent, not informed. They will produce a confident sentence containing a number that does not exist, and it will sit right next to a true sentence and look exactly the same.
Last year I asked for average open rates on post-purchase onboarding sequences. The answer came back with 47% and no source. The real figure across the accounts I could check sat closer to 38%, and it swung wildly by industry. Publishing the first number would have been a small, invisible lie.
So the rule is blunt: every statistic, date, name, quote, price, and product detail gets verified or replaced. If you can’t verify it, cut it.
The better move is to feed the model real material instead of asking it to remember. Paste in your own transcripts, customer interviews, support tickets. Working from AI meeting notes is a good example. The model isn’t inventing what a client said, it’s cleaning up what they actually said, and the output is dramatically more useful as a result.
Step 5: The Twenty-Minute De-Fluff Edit
Every model has tics. Learn yours and hunt them with search-and-destroy. Here’s my current list:
- “In today’s…” anything. Delete the opener entirely and start on sentence two.
- “It’s important to note that” or “it’s worth mentioning.” Say the thing instead.
- “Delve”, “leverage”, “robust”, “landscape”, “tapestry”, “testament to”.
- Rule-of-three lists in every single paragraph. Keep one, cut the rest.
- Em dashes used as a punctuation crutch. Swap most for periods.
- Paragraphs that end by restating the paragraph.
Read the final version out loud
This catches what the checklist misses. If you run out of breath mid-sentence, the sentence is too long. If your voice drops into a monotone, the paragraph is too uniform. If a line sounds like a brochure, cut it.
Step 6: Sort Out Disclosure and Detection
Two admin jobs people skip and then regret.
First, check the contract. Plenty of clients now require disclosure when AI touched a deliverable, and some forbid it outright for legal or brand reasons. Second, remember that AI detectors are unreliable in both directions. They flag human writing as machine-generated and let plenty of raw output sail through. If a client is judging your work by a detector score, knowing how AI-generated content actually reads is a far better defense than a percentage from a tool that guesses.
Where This Workflow Pays Off Fastest
Not every job deserves all six steps. The ones that reward it most:
- Volume work with a fixed format. Two hundred product descriptions from a spec sheet. The brief is identical every time, so your fifteen minutes amortises to almost nothing.
- Turning conversations into documents. Meetings, interviews, and sales calls become follow-up emails and case study drafts in a fraction of the time.
- First drafts of structure. Page layouts, section order, argument flow. A 90-minute AI website build is mostly this. The model proposes, you decide.
- Exploring directions you’d never try. The same way AI painting tools let artists test fifty compositions before committing to one, text models let you test angles before committing to a draft.
Where it doesn’t pay off: anything where the value is your specific judgement in the moment. Strategy memos. Sensitive client emails. Anything with your name on it that a client will read closely. Those you write yourself, and maybe use a model to check for typos.
Keep a Swipe File of Your Own Edits
Here’s the habit that compounds. Every time you fix a sentence the model wrote, paste the before and after into a running document. After two months you’ll have a personal list of the model’s tells for your specific subject matter, which beats any generic prompt guide.
My file is up to about ninety entries. The pattern in it surprised me. The fixes are rarely about accuracy. They’re about specificity. The model writes “improve customer retention.” I write “cut the unsubscribe rate from 1.4% to under 0.6% in one quarter.” Same idea, and only one of them is worth reading.
That gap between generic and specific is where your work lives. The generation part is now close to free. The judgement about which ninety seconds of output is worth keeping, that’s still the job, and no model has taken it yet.