Ask an AI generator for text on “email marketing” and it hands back 400 words of polished nothing. Every sentence is grammatical. None of it says anything you couldn’t find on the first page of search results. The tool isn’t the problem. The instruction was.
Most weak AI output traces back to a prompt that named a subject but never explained who’s reading, what the piece has to accomplish, or what good looks like. Below is the six-step workflow I use for client work, with the actual prompts. It adds about fifteen minutes on the front end and routinely saves an hour of rewriting.
Step 1: Write a five-line brief before you open the tool
Skip this and you’ll spend the next twenty minutes nudging the model toward something you should have specified at the start. Before typing a prompt, fill in five lines: audience, the job the piece has to do, format and length, tone limits, and the facts that must survive the edit.
For a landing-page hero at a meal-kit startup, that brief reads: busy parents who cook four nights a week; convince them to start a trial; three short paragraphs, 180 words total; warm but not cutesy; must mention the 20-minute average prep time and the free first box. That brief now becomes the spine of every prompt that follows.
Step 2: Feed it raw material, not adjectives
Telling a model to “sound expert” does almost nothing. Handing it the raw material it should be reasoning from changes everything. Paste your call notes, the last three emails that got replies, a competitor’s page you want to beat, the spec sheet, the customer support tickets.
A model working from 800 words of your own interview transcripts will produce copy that names the exact objection your buyers raise. One working from a one-line topic prompt will produce the generic version of that topic, every time. Context beats cleverness.
If you’d rather inspect the model yourself or run something locally, hubs such as Hugging Face host over a million open AI models you can browse directly, which is handy when you want to know what a writing tool is actually built on.
Step 3: Structure the prompt in four parts
A prompt that works consistently has four pieces:
- Role: who the model writes as. “You are a technical writer at a cybersecurity firm.”
- Task: one deliverable, described precisely. Not “write about backups” but “write a 200-word troubleshooting section for IT admins whose nightly backup job fails with a permissions error.”
- Constraints: length, reading level, banned words, formatting. Models obey explicit limits far better than vague ones.
- Example: one paragraph of writing that sounds like you. This does more for voice than any adjective you can name.
Assembled, a working prompt looks like this:
Sample prompt: “You are a B2B copywriter. Write a 90-word follow-up email to a prospect who downloaded our pricing guide and hasn’t replied in eight days. Direct tone, no exclamation marks, no ‘just checking in.’ Reference the ROI calculator on page 4. Here’s an example of the voice I want: [paste 60 words from a past email that worked].”
That’s roughly 70 words of instruction for a 90-word deliverable. The ratio feels absurd the first few times you do it. It isn’t.
Step 4: Get the skeleton before the prose
When you need 1,200 words instead of 90, don’t ask for the whole draft in one shot. Ask for an outline first: “Give me five section headings with one sentence describing what each covers, aimed at [audience].” Reordering five headings takes a minute. Reordering a finished draft takes twenty, and models tend to repeat themselves when you ask them to fix structure after the fact.
Once the outline holds up, expand one section at a time. “Write section 3 only, 250 words, using the interview quotes above.” You keep control of pacing and the output stays on the rails. Tools designed for this chunked approach, like the assistant covered in this review of HyperWrite’s writing workflow, let you park prompts and tone rules in a sidebar instead of retyping them all day.
Step 5: Rewrite the first and last sentence of every paragraph
This single habit removes most of the robotic texture from AI drafts. Models default to putting a topic sentence at the front of each paragraph and then restating it at the end. Read three of those paragraphs in a row and you hear the same rhythm, the same connective tissue, the same fondness for three-item lists.
Rewrite each opener so it lands differently. Delete the wrap-up sentence outright. Vary paragraph length on purpose. Then read the whole thing aloud, which is the fastest test for whether a sentence was written or generated. The sentence-level fixes in this guide to removing the robotic tone from AI writing go deeper on exactly that.
Speed isn’t the asset here. Your voice is, and there’s a solid case for protecting it in how to get real value from AI text generators without losing your voice.
Step 6: Check every number, name, and quote
These systems predict plausible text. Plausible and true are different things. They will invent a statistic with total confidence, attribute a quote to the wrong person, and cite a study that doesn’t exist.
A fast pass: highlight every number, date, proper noun, and quotation in the draft, then verify each one against a primary source. On a typical 1,000-word draft that takes under ten minutes and catches something nearly every time.
Keep the prompts that worked
The compounding win isn’t any single draft. It’s the folder of prompts that reliably produce a usable first pass. Save each one with a note: “cold email, SaaS, 90 words, worked 4 out of 5 times.” Six months in, that folder is worth more than any subscription.
Before anything goes live, run it against this list:
- Does the opening sentence say something specific, or does it just warm up?
- Could a competitor publish this paragraph unchanged? If yes, it’s too generic.
- Is every statistic traceable to a source you have actually read?
- Would you say this sentence out loud to a customer with a straight face?
- Does it contain one detail only you could know?
That last item is the dividing line. An AI generator for text can deliver competent structure, clean grammar, and a fast first pass at almost anything you throw at it. The detail only you have is what makes the result worth reading.
Once a draft clears that list, it can travel. The same 300-word script that works as a blog section usually works as a video voiceover with light trimming, and tools like InVideo AI turn a text prompt into a finished video with captions and b-roll in a few minutes. Write once, then decide how many shapes it takes.