An AI Workflow for Instagram Content That Doesn't Sound Like AI

AI is now good at several parts of running an Instagram account and actively harmful at one of them. Knowing which is which is most of the skill.

Where AI genuinely helps

1. Idea expansion

Give it your audience, your topic and a constraint, and ask for twenty angles. It will produce nineteen bad ones and one you would not have thought of. That is a good return.

2. Drafting hooks

Hooks are short, structural and formulaic, which is exactly the shape of problem a language model handles well. Ask for twenty variations of a hook for a specific video and pick by instinct.

3. Editing your own draft

Paste what you wrote and ask it to cut 30% without losing meaning, or to rewrite a section in a plainer register. Editing is where AI is most reliable, because it is not inventing anything.

4. Turning one asset into another

A Reel script into a carousel outline. A carousel into a caption. A caption into a story sequence. Mechanical work, well suited.

5. Alt text

Ask for a description of an image; you will need to correct it, and it is faster than starting from nothing.

6. Analysing your own numbers

Paste a month of post metrics and ask what the outliers have in common. It is pattern-spotting on a spreadsheet, and it is often right.

Where AI hurts

Generating the opinion. A model has no experience of your niche, your tests or your audience. Ask it to write "what actually works on Instagram" and you get the average of the internet: confident, generic, and identical to a thousand other accounts. The one thing that makes an account worth following — a specific, earned point of view — is the one thing it cannot supply.

Producing statistics. It will invent them, in a believable format, if you let it. Never publish a number you have not seen in your own insights or a primary source.

Writing the first draft of a whole post. The output reads like a summary because it is one.

A prompt structure that avoids the AI voice

The problem in most AI output is not the model; it is the missing context. Four things to include:

1. Who is speaking. "I run an Instagram growth blog. I have tested this on accounts between 500 and 50,000 followers."
2. Who is reading. "The reader has under 2,000 followers, posts three times a week, and has read all the standard advice already."
3. The claim. State the specific, arguable position the piece takes. If you cannot state it in one line, you do not have a piece.
4. The constraints. No emoji bullets. No "in today's digital landscape". No invented statistics. Short sentences. British or American spelling, pick one.

Then ask it to structure that claim rather than invent one.

A workflow that keeps the voice yours

1. You write the one-line claim. This is the non-negotiable human step.
2. AI expands it into an outline of five to seven points.
3. You delete the points that are wrong and keep the ones that surprise you.
4. AI drafts one paragraph per surviving point.
5. You rewrite the opening and the closing yourself — those are the parts readers remember.
6. AI cuts 20% for length.
7. You check every factual claim.

Steps one, three, five and seven are yours. The other three are assembly, and assembly is what you are paying for.

The disclosure question

If a piece is substantially written by a model from someone else's reporting, say so. Our own editorial standards describe the rule we hold ourselves to: AI is a tool for writing, not a source of facts or experience, and any piece that begins as another publication's work carries a visible attribution.