An AI copywriting framework stops generic copy by making the model answer four questions before it writes a line: who are we talking to (persona), what bugs them (tension), why it happens (angle) and what one line says it (story). When each answer comes from real posts and audience data, the copy has something specific to say, and the model stops reaching for the average.
Key takeaways
- Generic AI copy is mostly a briefing problem: a model given no reader, no problem and no reason writes for everyone.
- The four questions are persona, tension, angle and story, and each one narrows the space the model can drift into.
- The angle is the step most teams skip; naming the behavioural theory behind it, such as Overconfidence Bias or Self-Determination Theory, makes the copy argue something.
- Tie every answer to evidence, then rewrite the lines that fail the audience check, as SOMIN did for Keeta HK.
Why does AI copy sound generic?
Because a language model with no specific reader or problem writes towards the most likely sentence, and the most likely sentence is the one your competitors get too. The fix is a sharper brief, not a cleverer model.
Research points the same way. A study by Anil Doshi and Oliver Hauser in Science Advances, summarised by the University of Exeter, asked 300 people to write short stories, some with AI-generated ideas. Writers given AI ideas produced stories that evaluators rated as more novel and more useful, yet the AI-assisted stories were more similar to each other. Each writer improved while the group converged.
That is the risk for a category in which every brand uses the same tools. SOMIN's ads-optimisation workspace names it in an insight, "Authenticity is the antidote to homogenization", with an angle of Cultural Homogenisation: brands lean on similar AI templates and formats and lose a voice of their own. SOMIN's moments-of-truth model surfaces the matching thought in the audience-research workspace: "If we don't find a unique angle, we risk becoming invisible to our audience."
Anthropic's own prompting guide gives a test that applies to any brief. Its golden rule is to show your prompt to a colleague with minimal context; if they would be confused, Claude will be too. "Write a post about our new feature" fails that test every time.
What are the four questions?
Who are we talking to, what is bugging them, why does it happen, and what is the one line that says it. In SOMIN's language those are persona, tension, angle and story.
| Question | Field | A strong answer | A weak answer |
|---|---|---|---|
| Who are we talking to? | Persona | A segment with a role, goals and context, drawn from the people who engage with category content | "Marketers aged 25 to 45" |
| What is bugging them? | Tension | A specific friction, described as they would describe it, with the number of posts that carry it | "They want better results" |
| Why does it happen? | Angle | The mechanism behind the tension, named by a behavioural theory | Missing, or "because it is hard" |
| What is the one line that says it? | Story | One sentence the persona would recognise and only this brand could say | A tagline any competitor could run |
The order matters. A tension without a persona is a complaint about the world; an angle without a tension is a theory looking for a problem; a story without an angle is a slogan.
How does the framework work on real data?
The clearest way to see it is to hold the persona steady and change the angle. These three insights come from SOMIN's social-listening workspace, where agency and in-house social teams are the audience.
Overconfidence Bias: the multi-channel launch
Tension: brands overestimate how many platforms they can run well, and a one-size-fits-all multi-channel launch falls flat. Angle: Overconfidence Bias, from behavioural economics: the tendency to overrate one's own ability and control. The insight, drawn from 69 posts, reads "Overreaching in confidence often leads to hidden pitfalls." Copy built on this angle should sell focus and honest scoping, not more channels. An illustrative line: "Three channels done well beat seven done in a hurry."
Self-Determination Theory: the polished-post problem
Persona: the Dynamic Digital Strategist, an agency social lead. Tension: polished, perfect posts underperform raw, real stories. Angle: the desire for authenticity, read through Self-Determination Theory, which says people are driven by autonomy, competence and relatedness, so a staged, flawless post offers little to relate to. The insight, "Authenticity thrives in the cracks of perfection", sits on 330 engagements across 24 posts. Copy here should show the work in progress and the person behind it.
Symbolic Interactionism: when automation feels cold
Tension: automation makes brand interactions feel transactional. Angle: Symbolic Interactionism, which says meaning is built in the exchange itself, so a canned reply signals that nobody is listening. The insight, "Authenticity thrives when technology takes a backseat to human connection", carries 488 engagements across 36 posts. Copy should put the human in the loop in plain view.
Same audience, three angles, three different pieces of copy. Skip the angle and all three collapse into "authentic content that resonates", which is precisely the generic line we set out to avoid.
What does this look like on a live brand?
The Keeta HK case study runs the full loop and shows why the story step needs an evidence check. SOMIN benchmarked 145 ads from Keeta HK, FoodPanda and Deliveroo and found Keeta trailing both rivals on the category's dominant tension: affordable, convenient food.
The focus persona, the Flavourful Culinary Enthusiast, was validated against the same ad data. SOMONITOR AI generated four stories, each in three tones, and the team checked every line against a GWI snapshot of Hong Kong food orderers in the above-average income bracket. That audience under-indexed on budget-driven behaviour and over-indexed on curated, quality-first choices.
So the raw line "Quick to your door, kind to your wallet—Keeta makes every bite worth it." became "Smart meals for smart people—Keeta delivers fast, curated, and made for you." Across the set, 36 AI-generated lines in English and Cantonese were matched by 36 audience-validated rewrites. The persona and stories stayed; the price language went.
How do you brief an AI tool with the framework?
Put the four answers, their evidence and what to avoid into the prompt, then ask for several lines per tone rather than one perfect line.
- Persona. One paragraph on who they are, taken from data rather than a demographic guess.
- Tension. The friction as the audience describes it, and how many posts carry it.
- Angle. The theory by name and one sentence on the mechanism.
- Story. Ask for three lines in each of three tones: emotional, direct-to-action and playful.
- Exclusions. List the category's dominant message and any claim a competitor already owns.
- Check. Strike any line that a competitor could run unchanged, and test the rest against audience data.
SOMIN's Strategy agent builds these chains from live category data, and the Insight Engine skill in the free Marketing Mondays plugin lays out persona, tension, angle and story chains with the posts behind each one. Whatever tool you use, check the output before it ships; our QA checklist for AI marketing content covers what to look for.
FAQ
What is an AI copywriting framework?
It is a fixed set of questions you answer before an AI tool writes copy, so the model works from a specific reader, problem and reason. Persona, tension, angle and story is one such framework.
What is the difference between a tension and an angle?
The tension is what bothers the audience; the angle explains why it happens. Two brands can address the same tension from different angles and end up with very different copy.
Why use behavioural theories in copywriting?
A named theory, such as Overconfidence Bias, describes a mechanism the copy can work with. It gives the writer, or the model, a reason for the line instead of a mood.
Does the framework work with ChatGPT, Claude or Gemini?
Yes. The framework sits in the brief, so it works with any model. The quality depends on whether the four answers come from evidence or from guesswork.
Pick one live brief this week and rewrite it as four short answers, each with the evidence behind it. Ask the model for nine lines across three tones, strike every line a competitor could run, and test the survivors against what your audience data says.
