Drafting Social Media Copy With AI: The Editing Workflow That Keeps It Human
A round-by-round prompting and editing process for turning AI social drafts into posts that sound like your brand, plus the exact phrases that give away a machine wrote them.
The first draft isn't the problem, treating it as final is
Ask a model to write an Instagram caption and you'll have something in about four seconds. You'll also have something that reads like it was approved by a committee that has never once talked to your customers. That's not a reason to skip AI for social copy. It's a reason to treat the first output as a rough cut instead of a finished post.
More small teams are leaning on AI drafts to keep daily posting schedules alive, especially accounts with no dedicated copywriter. The trouble shows up after the draft lands, in the gap between technically correct and actually sounds like us. Closing that gap is a skill, and most of the skill is knowing exactly what to cut.
A five-round workflow, tested on a real product launch
- 1Round 1: get the ugly draft out fast
A four-person coffee roaster in Tacoma needed a caption for a new Ethiopian single-origin bean. The first prompt was bare: 'write an Instagram post announcing our new coffee.' The output opened with 'Attention coffee lovers!' and closed with 'Don't miss out — order yours today!' Useless as a final post, but useful as scaffolding.
- 2Round 2: feed it the specifics only you know
The second prompt added real detail: the farm name, the altitude, the tasting notes the roaster actually wrote on the bag (blackberry, brown sugar, a little like Earl Grey), and the price. Specifics do more to kill generic tone than any instruction about 'voice' ever will.
- 3Round 3: name the tells and cut them
The draft still had 'Are you looking for a bold new coffee experience?' and three exclamation points in two sentences. Both got cut with a direct instruction: no rhetorical questions, no more than one exclamation point per post, no em-dash before a call to action.
- 4Round 4: add a stake or an opinion
AI copy tends to sit on the fence — it describes, it doesn't take a side. The roaster added one line from the person who actually roasted the batch: 'I ran this batch four times before the roast curve was right.' That single sentence did more for authenticity than any adjective swap.
- 5Round 5: read it out loud, then trim for the platform
The final pass was a human one: read the caption aloud, put back the contractions AI drafts tend to drop, and shorten it for Instagram versus the longer version kept for the shop's email list.
Six tells that flag a caption as AI-written
- Rhetorical questions as an opener: 'Ever wondered why your coffee tastes flat?'
- No contractions anywhere — 'it is' instead of 'it's,' 'we are' instead of 'we're'
- A closing line that restates the whole post in different words
- Stacked adjectives in threes: 'bold, rich, unforgettable'
- A call to action that could apply to literally any product: 'Don't miss out — shop now!'
- Hashtags that read like keyword stuffing rather than something a person would actually type
Platform differences, briefly
| Platform | One adjustment worth making |
|---|---|
| Trim the AI draft by roughly a third — it tends to over-explain | |
| Keep the industry specifics AI drops in favor of generalities | |
| X | Cut the draft to one idea; AI drafts often try to fit three |
AI models will confidently invent a price, misstate a launch date, or attribute a quote to the wrong team member if your prompt was vague. Treat every number, name, and claim in a draft as unverified until someone on your team checks it against the source — a pricing sheet, a calendar, an actual person.
Why iteration beats a longer first prompt
It's tempting to try solving all of this up front with one exhaustive prompt: describe the brand voice, list the banned words, specify the tone, all in one shot. In practice, a model handed twelve instructions at once tends to satisfy about half of them and quietly drop the rest, especially the ones buried in the middle of a long prompt. Short, sequential rounds work better because each one only has to solve a single problem — the coffee roaster's five-round process worked precisely because round three only had to fix rhetorical questions and exclamation points, not the whole caption at once.
This also matches how editing actually works when a person does it alone, without AI involved. Nobody rewrites a paragraph for structure, tone, and grammar in the same pass — they read it once for one thing, then again for another. Treating an AI draft the same way, as a rough starting point that gets refined in stages rather than accepted or rejected as a whole, produces noticeably better captions than judging the first output against a finished-post standard it was never going to meet.
Keep a one-page voice reference next to the prompt
The fastest way to stop re-fighting the same battles every time you draft is to write down what you keep cutting. A one-page doc with three real captions you're proud of, five phrases you've banned, and the sentence length your account actually reads at saves more editing time than any clever prompt trick.
Paste that doc into the prompt itself. Models follow concrete examples far better than abstract instructions like 'sound more casual.'
If you want a second opinion on whether a caption still reads as machine-written after your edits, run it through AI Humanizer Lab's free AI detector — it flags the sentences most likely to trip up a reader, and the humanizer tool can smooth over the ones you'd rather rewrite by hand than delete outright.
Make your writing sound human
Humanize AI-generated text in one click with AI Humanizer Lab.
Try for free