How to Write UGC Ad Scripts With ChatGPT That Don't Sound Like ChatGPT

Before you open ChatGPT, pull ten to fifteen sentences from reviews, tickets, and post-purchase surveys where customers name the problem or the result in their own words. In the same note, list five to ten phrases your brand must never say on camera — superlatives, clinical claims, and the marketing verbs you already know sound fake. That packet is the brief. ChatGPT mirrors whatever you feed it; feed it homepage copy and you get a homepage with a hoodie.

Prefer phrases that repeat across customers over one colorful outlier. If six reviewers independently write some version of 'I did not think it would work this fast,' that skepticism-to-surprise arc is a hook, not a testimonial to paste. Copy the wording, including the awkward bits. The awkward bits are why it sounds human. Brand-written benefit lines like 'designed to elevate your routine' are the opposite input: they teach the model the cadence you are trying to escape, and ChatGPT will reproduce that cadence at scale if it is the only prose in the packet.

The banned list is as important as the source quotes. Put 'game-changer,' 'unlock,' 'elevate,' 'seamless,' 'revolutionary,' and any unsubstantiated 'clinically proven' line on it before generation, not after. ChatGPT will reach for those words unless you forbid them in the same message as the customer quotes. Keep the packet in a running doc so every future script starts from the same raw language instead of a blank chat, and add a new ban whenever a draft sneaks a word past you. Put the date of the last ban at the top so the packet does not go stale.

Tell ChatGPT the script must hit four beats in order: a hook in the first two seconds, a problem-agitation beat that uses customer language, a demo beat that shows the product working, and a specific call to action. Ask for one body and three hook variants, with the middle and close held fixed. 'Write me a viral UGC script' is how you get a press release. A beat sheet is how you get something an avatar can actually say.

State the job of each beat in the prompt so the model cannot collapse them. The hook earns the next three seconds and does not need the product name. Problem-agitation should quote or paraphrase the customer lines you pasted. Demo should be visual and checkable — 'watch what happens when I twist this open' — not 'engineered for effortless use.' The CTA names a next step ('tap the link and grab the bundle'), not 'shop now' with a smile. If you do not name the jobs, ChatGPT will smear benefits across all four beats and the cut will feel like one long pitch.

Give hard constraints in the same message: spoken first person, contractions required, one idea per sentence, target 75 to 95 words for a 30-second ad, no hashtags, no emoji, no stage directions. Paste the banned-phrase list. Paste two or three of the strongest customer quotes and tell it those phrases must appear, lightly adapted, in the problem beat. Models follow stacked constraints better than a mood ('make it casual and gen-z'). Casual is not a spec. Word count, beat order, and forbidden vocabulary are.

Add a rewrite pass whose only job is to make the draft sayable. Instruct ChatGPT to split comma-heavy sentences, add contractions, kill transition words like 'moreover' and 'that's why,' and replace any line you would not text a friend. Then read the result out loud at talking speed. If you run out of breath or stumble, the line is still a paragraph pretending to be speech.

ChatGPT defaults to written English with neat logical glue. UGC dies under that glue. A useful rewrite prompt is: 'Rewrite for a person talking to a phone camera. No sentence longer than sixteen words. Use contractions. Cut every transition that exists only to sound structured. If a line could appear on a product page, rewrite it as an experience.' Then you still have to speak it. The read-aloud pass catches what the model cannot hear: stacked clauses, claims that sound illegal out loud, and jokes that only work in text.

Time the spoken draft. Seventy-five to ninety-five words is the 30-second window that leaves room for pauses; longer scripts force a rushed delivery that undercuts the unscripted feel. If you are over, cut the problem-agitation beat first — it is the most tolerant of trimming. If a sentence is hard to say twice, delete it instead of polishing it. Difficulty speaking is usually a sign the line is doing two jobs, which is a brochure move, not a talking-head move, and ChatGPT loves stacking jobs in one clause.

Do not ask for 'a script with a great hook.' Lock the body, then run a second prompt that produces twenty hooks in labeled formulas: question, confession, complaint, number, and pattern interrupt. Each hook must be ten words or fewer and must not name the product unless the formula is a direct callout. The hook is the test variable. Mixing a new hook into a new body every time makes the later test unreadable.

Give each formula a one-line definition in the prompt so ChatGPT cannot collapse them into paraphrases of the same announcement. Question hooks ask something the viewer already feels. Confession hooks admit a late discovery. Complaint hooks name a hated workaround. Number hooks put a specific, true figure in the first breath. Pattern interrupts start mid-action. Ask for four of each, then throw out anything that could run for a competitor with the product name swapped. Specificity is the filter: if the line still works for a rival SKU, it is not a hook yet.

Ban the model's favorite openings in this prompt: 'Are you tired of…,' 'Imagine a world where…,' 'What if I told you…,' and 'Stop scrolling.' Those are ChatGPT's default UGC costume. Require at least one concrete noun from your customer packet in every surviving hook. A confession that names the exact failed product or the exact messy moment will outperform a generic 'I cannot believe I waited this long' every time you test, because the viewer can picture it before the face even finishes the sentence.

Read the full script and highlight every line that sounds like a model: perfectly balanced clauses, fake enthusiasm, sudden lists of three benefits, and any sentence that explains the ad to the viewer. Rewrite those in the customer's vocabulary from step one, or cut them. The goal is not prettier copy. The goal is a script that would not get you clocked as AI if a creator posted it from their bedroom.

Common tells: 'that's the thing,' 'here's the best part,' 'and the best part is,' benefit tricolons, and a CTA that restates the entire pitch. Also watch for unearned intimacy ('trust me') and unearned expertise ('as someone who has tried everything'). Replace those with a single specific result or a single visible action. If the demo beat lists features, turn the list into one action the camera can show. A talking head reciting three bullets is a landing page. A talking head twisting a cap is a UGC ad.

Do a mute test on the page: if you cover the first line, does the second line still sound like a person picking up a thought, or like slide two of a pitch deck? UGC should feel like you started recording after the conversation had already begun. That usually means cutting the warmup sentence ChatGPT adds to 'set context.' Context is the viewer's problem, not your first two seconds. Save the product name for after the hook has earned it, unless the hook formula is a direct product callout you are deliberately testing.

Go line by line and flag results, speed, health, income, and comparisons. For each flagged line, either attach the evidence you actually have or rewrite it as a first-person experience with a hedge. ChatGPT is fluent at sounding certain. Platforms and lawyers are not impressed. An ad that dies in review wastes the whole generation sitting, and a softened line still sells if the demo is visible.

Prefer experience language over authority language. 'I noticed less redness after a few days' survives review more reliably than 'clinically proven to reduce redness,' even when a study exists, because the spoken line is not the place to litigate a claim. Keep a one-page approved-claims list and paste it into the ChatGPT thread when you ask for rewrites so the model substitutes from the list instead of inventing a new proof point for every variant. Invented proof is the most expensive kind of fluency.

Check on-screen captions and text overlays too. It is common to soften the voice line and then let a generated caption restore the absolute. If you will burn in captions, write them in the same pass as the script, not as an afterthought from another ChatGPT prompt that never saw the claims list. Consistency across voice and type is how you avoid a rejection that looks like a creative problem but is actually a compliance problem you already solved once and then undid in a second chat. Read captions last, after the voice is locked.

Break the script into short delivery lines with a slash at every natural pause, mark emphasis, and attach the three hook variants as alternatives, not as new bodies. Then generate or film the video against that locked page. A ChatGPT draft is not a script until it has line breaks a mouth can follow and a version number you can find six months later when the ad is still spending.

If an AI avatar will deliver it, those line breaks also drive pacing, so a script formatted for breath becomes a more natural render. If a human creator will film it, send the formatted page plus permission to swap a word in the hook so it fits their mouth. Word-for-word recitation is how ChatGPT scripts get clocked even after you cleaned the vocabulary. Small deviations are the unscripted feel. Keep the locked text in the same folder as the export so you can clone winners without guessing which chat produced the line.

Klip Kanvas can take the locked four-beat script and render talking-head variants across avatars while you keep the body fixed and swap only the hook lines from the separate generation pass. That is the point of splitting hooks from body in ChatGPT: production should not reinvent the middle. Watch the first render against the page, confirm the hook lands in two seconds, and do one more caption pass so a claim you killed in text did not return on screen because the renderer paraphrased it. If it did, fix the caption file, not the spoken take.