Best Prompt Frameworks for UGC Ad Scripts, Ranked

The prompt that consistently survives a buyer's edit is the one that looks like a production brief. You state the SKU, the one mechanism, the one allowed proof, a banned-word list, the structure with timestamps, the spoken-word cap, and a table schema for the reply. The model is not asked to be creative about claims. It is asked to arrange what you already believe is true into a 20–35 second talk-track a person could say on a phone.

Minimum blocks: product facts (only these numbers exist), audience in one sentence, structure name, beat map in seconds, first-frame action, CTA formula, don'ts (no 'honestly', no 'wait for it', no invented reviews), output as table columns — time, visual, spoken, caption. Require a claims log at the bottom mapping every factual clause to a line you pasted. If a number is not in the brief, the cell must be empty, not guessed. That single rule kills most of the confident nonsense. Keep the schema identical week to week so a buyer can diff two runs. If the model returns a paragraph instead of the table, reject the run and resend the schema only — do not chat it into shape.

Reuse the same skeleton every week so scripts are comparable. When you generate video from the output, including in Klip Kanvas, paste the don'ts and the first-frame line into the product brief rather than hoping the avatar tool infers them. The framework fails when the constraint list is a novel — six pages of brand voice will be ignored. Keep don'ts under fifteen bullets. Keep facts under ten. If you cannot fit the truth on one screen, you do not have a UGC ad yet; you have a PDP. When a founder wants to add brand values, put them in a single 'tone' line, not a manifesto. Values that cannot fit in one line are not a UGC constraint; they are a website footer.

This framework does not ask for 'some variations'. It asks for a labelled grid. You paste a frozen body (beats 8–25s and the CTA), name two hook formulas and two angles, and require four complete opens that do not share a first three words. The rest of the ad stays still. That is how a later hook-rate column can be trusted. Without the freeze, the model will 'improve' the middle of every cell and you will have four different ads.

Name the formulas in the prompt — pattern interrupt, review-read, problem, demo-first — instead of 'give me hooks'. Paste one example of each formula that already worked for you, or a stolen structure without the competitor's claim. Demand a first-frame note per cell (hands, pack, face, text). Ban synonyms of the same open; 'okay so' and 'so basically' are one hook. Output must include a cell ID you will put in the file name. Number the cells A1–B2 in the prompt and in the file name. If a hook formula has no example, the model will write a rhetorical question and you will think you tested a pattern interrupt.

After you have a winning row, the next prompt is not a blank page. It is 'keep angle A, replace the two hooks, freeze body'. That inheritance is the whole point of ranking this second. If production cannot shoot or generate four modular opens this week, do not run the prompt; you will get a document. Pair with a generator that can fan out hook × avatar so the matrix is footage, not a Notion table. Archive the frozen body in the same doc as the prompt so next week's writer cannot 'just tweak the CTA'. Inheritance only works if the body actually stayed still.

You paste one real review (or three) and a structure: read the line, look up, prove it. The prompt's only creative job is the proof beats and the afterthought CTA. Ranked here because the first three seconds arrive pre-tested by a customer. The model is a threat to this framework: it will sand off the rudeness that made the line work. Your system prompt has to say keep the customer's nouns and numbers.

Quote the review inside delimiters and say 'do not paraphrase the quoted sentence'. Ask for a sceptical delivery note, not a grateful one. The proof that follows must test the claim in the quote — if they said 'didn't pill after three washes', the B-roll is fabric, not a smile. Attribute vaguely if you need to ('someone wrote') but never invent a name. If the review implies a health outcome you cannot claim, reject it in the prompt instructions rather than rewriting it into a miracle the legal team will pull.

Batch it: three reviews in, three scripts out, one claims check. This is the fastest path from a live PDP to AI UGC that still sounds like UGC. It collapses on new listings. Do not ask the model to 'imagine a review'. That is fake testimonial territory, which is a worse problem than a weak hook. Wait, or use PAS until the page has a sentence worth stealing. When three reviews share the same claim, pick the rudest specific one and kill the other two as cells — they are duplicates, not a matrix. Save the extras for fatigue week, not for launch day.

Demo-first ads die in language models because text-only prompts privilege speech. This framework forces two passes: first a silent shot list with timestamps and what the hands are doing; second a voiceover that may only describe what is already on the list. The first frame becomes a verb. Muted play still holds. It is the right prompt for kitchen, cleaning, beauty texture, gadgets — anything whose mechanism is visible in three seconds.

Pass one output: 0.0–2.5s, 2.5–6s, and so on, each with camera height, object, action, no dialogue. Ban 'talking head' in the first two shots. Pass two: VO lines mapped to those IDs, plus a caption line that works with sound off. If a VO line has no shot, delete the line, do not invent a shot. This split also makes B-roll briefs usable in generators that take visual direction separately from script. If the generator only accepts one script field, paste the shot list into the visual brief and the VO into the talk-track field separately. Merging them in a single box is how you get a face again.

Where it fails: supplements, SaaS, and any SKU with nothing honest to show. Forcing a demo of a capsule bottle is how you get a hand holding a pack and a lecture. Use PAS or objection-then-proof instead. Also fail: merging the two passes in one prompt 'for speed'. The model will write a face-first script and decorate it with 'B-roll of product'. That is not a split. Make it two messages. Time-box pass one to ten minutes. Teams that 'just write the VO first this once' are the reason demo-first ads still open on a greeting. Make the split a checklist item, not a preference.

You paste the three reasons people bounce — price, 'does it work', shipping, fit — and require the script to voice the strongest one before the pitch. The model is good at this if you give it the real objections from comments and 1-stars. It is bad at this if you let it invent 'are you tired of low quality'. Ranked fifth because it is a specialist: cold traffic with a known stall, not a default for every SKU.

Structure the prompt as: objection (quoted), why the current workaround fails (one line), proof (demonstrable), close. Ban stacking all three objections into one 35-second ad; that is a FAQ. One doubt per cell. If price is the objection, the prompt must include how you are allowed to present the number — comparison, per-use, or not at all — or the model will blurt a discount you do not run. Pull the three bounce reasons from actual PDP analytics or from comment search, not from a brainstorm. If you cannot quote the objection, you do not have this framework yet — you have PAS with extra steps.

Pair with a sceptical avatar or creator note. A cheerful read of 'I know it's expensive' is an ad. A dry read is UGC. Put delivery energy in the prompt as a constraint, not as 'make it authentic'. Authenticity is not a setting. For AI UGC, this framework plus a doubt-friendly archetype usually beats a generic enthusiast on hook rate when the category is crowded with smilers. If the objection is shipping, the proof has to be a day-count you actually hit, shown on a packing slip or a calendar, not 'fast delivery' in a confident voice. Unprovable objections do not belong in this prompt.

Language models default to advertised-sounding copy because that is what 'write an ad' means in the training mix. This framework pastes a bad script — your last rejected take, or a parody of UGC tropes — and lists the moves that made it an ad: smile on the logo, 'I have to share this', music swell, feature list. The new draft may not reuse those moves. It is a complement to constraint-first, not a replacement for facts.

Be specific in the autopsy. 'Too salesy' teaches nothing. 'Opened on a greeting, used the word honest, never showed the mechanism, closed on SHOP NOW shouted' is a lesson. Ask the model to label each forbidden move before writing, so you can see whether it understood. Then still freeze facts and structure; anti-ad alone will give you a vague vlog. Paste the rejected take in full, including captions. The tropes often live on screen, not in the VO, and a text-only autopsy will miss them. Ask the model to list forbidden on-screen text as its own row.

Rotate the negative example. If you always paste the same parody, the model over-indexes on avoiding those five phrases and invents a new set of tropes ('let me be real with you'). Rebuild the don'ts from actual rejected takes once a month. This framework is especially useful when you are generating at volume and the library is starting to sound like one person. Keep a running 'graveyard' of five rejected opens and rotate which one you paste. That stops the model from overfitting to a single parody and writing the same anti-ad every time.

Instead of one prompt that writes the whole spot, you generate ten first lines, pick one, then generate the body against that line, then generate only the close. Each stage is a human gate. Ranked seventh because it is slower and it still fails if the picker has no taste — but it is the best recovery when a one-shot brief keeps missing the first three seconds. The chain is a process, not a magic prefix.

Stage one: twenty hook lines, no body, labelled by formula. Kill fourteen in a minute. Stage two: body beats that may not restate the hook. Stage three: afterthought CTA, not a new pitch. If you let the model 'polish the whole script' at the end, you undid the gates. Keep the artefacts: the chosen hook ID lives in the file name so the test still isolates something. Limit stage one to a table of lines, not monologues. If a hook needs a second sentence to make sense, it is already a body. Kill it at the gate instead of hoping the next stage will rescue it.

This is a poor default for a solo founder who needs ten ads by Thursday — use constraint-first plus a matrix instead. It shines when you have a working body and a dead hook, or when legal must sign the proof lines separately from the open. Do not confuse it with chain-of-thought padding ('think step by step') which often just produces a longer preamble and the same mediocre hook. Assign one person as picker; a committee at each stage is how this framework becomes a workshop. The chain is cheaper than a shoot only if the gates take minutes, not a meeting.

Persona prompts feel like strategy and produce adjectives. Age, city, job title and 'relatable tone' do not constrain claims, do not specify a first frame, and do not stop the model from writing a compliment with a URL. Casting matters — avatar archetype and real creator fit move hook rate — but that is a production choice, not a substitute for a brief. Ranked last because it is the default in most ChatGPT ad threads we see, and the default is why the ads sound the same.

If you want a sceptical nurse, put that in one line inside a constraint-first brief, then spend the rest of the prompt on facts, don'ts and beats. Do not build a backstory. Models will spend tokens on the roommate and the morning routine and forget the SKU. Identity angles can be real (fit, profession, pet parent) when they are the angle axis of a matrix, not when they are a costume the copy is wearing. If a stakeholder insists on a persona paragraph, move it under the beat map so it cannot eat the facts. Then count how often the output still names the SKU in sentence one. That count is the argument against the dump.

The salvage: take any persona prompt you already like and wrap it in ranks 1–3. Keep one sentence of voice. Delete the biography. Add the claims fence. If the script still works, the persona was never doing the work. If it collapses, you did not have a product argument — which is information, and cheaper than finding out in the auction. Casting still belongs in production: pick the avatar or creator after the script exists, as its own test cell. Putting the biography in the prompt is how you confuse a media variable with a copy variable.