TemplatesScriptsHooksCheatsheet

ChatGPT Prompt Library for UGC Ads: Hooks, Scripts and QA

Copy-paste ChatGPT prompts for UGC ads: input pack, hook batches, timed scripts, variation grids and a claim-safe rewrite pass — with output schemas you can brief or render from.

Updated 2026-08-2412 min read

LanguageEnglish

A blank ChatGPT thread is not a creative workflow. This library is the prompts we paste in order: input pack, hook batch, timed script, variation grid, then a claim-safe rewrite — so the model fills structure instead of inventing a new ad format every time.

The model is a clerk, not a creative director

ChatGPT will happily write a 90-second manifesto with a new brand voice every turn if you ask it to 'write a UGC ad.' That output is not testable: you cannot compare five of them, you cannot time them, and you cannot tell whether the hook or the offer moved. Lock the job in the system block, lock the skeleton in the user block, and lock the output as labelled fields. Your job is the input pack — product URL facts, real review phrases, the offer that is actually on the landing page, and the claims you will not make. The model's job is volume inside that fence. If a prompt could be used on any product without those inputs, it will return generic ads that die on cold traffic.

Run prompts in a fixed order or you will rewrite the same ad

Order is the workflow: (1) pack the facts, (2) generate hooks only, (3) expand one hook into a timed script, (4) spin variations on one variable, (5) run the QA rewrite. Jumping to 'write 10 full ads' mixes hook, body, CTA and claims in one blob, and you cannot kill the weak half. One ChatGPT project or custom GPT should hold the system block and the banned-claim list; each new SKU is a new user pack, not a new personality. Save winning hooks as a list you paste back in — do not ask the model to remember last week's winners from chat history.

What to do with the output

Treat every reply as a draft table, not a file you upload. Time the lines out loud; if a 30s script reads past 85 words, it will not fit a talking-head. Check the landing-page offer against the CTA. Then generate the actual video from the beat list, not from the chat transcript. Klip Kanvas will take a product URL plus a beat-mapped script and return 9:16 UGC; ChatGPT's job ended when the fields were filled. If you paste raw ChatGPT prose into a renderer without timings, you will get a voiceover essay with no hook.

How to use the tables below

Copy the system block into a ChatGPT project once. For each SKU, paste the input-pack template with real URL facts and quoted review phrases, then the hook-batch prompt, then one timed-script prompt for the three hooks you keep. Run the 3×2 variation prompt only after a body is locked, and the claim-safe QA on the whole grid before anyone renders. Locked rows are the literal prompt skeleton and the required output fields — not essays about prompting. Swap bracketed slots; do not delete the constraints or the model will invent a new ad format. If it ignores the schema, resend the output-format row only. Time every script out loud: over 90 spoken words is a fail, not a 'punchy' read.

5 prompt tables inside: system block plus input pack, hook-batch prompt with CSV schema, 30s timed-script prompt, 3×2 variation prompt, and claim-safe QA rewrite with fail codes and required output fields.

System block and input pack (paste first)

System role: "You write short-form UGC ad copy for paid social. You fill templates. You do not invent ingredients, results, reviews, prices or guarantees. If a fact is missing, write NEED: [field]."
Hard limits: Talking-head 15–35s · hook in first sentence · no 'game-changer' / 'holy grail' / 'doctors hate' · no diagnosis of the viewer · UK/US English as specified
Input pack — product: URL · SKU name · 1-sentence mechanism · 3 proof points you can substantiate · 3 exact customer phrases from reviews (quoted)
Input pack — offer: Price · discount code · shipping · guarantee — only if live on the PDP · else omit
Input pack — audience: Cold / warm / retarget · platform (Meta Reels / TikTok / Shorts) · language · banned claims list
Input pack — identity: Peer UGC vs founder vs demo-hands · do not let a peer say 'I founded'
Refusal rule: If proof list is empty, output hooks that do not assert results · never fabricate a number

Hook-batch prompt (hooks only)

User prompt stem: "Using ONLY the input pack, write 12 hooks. 8–16 words each. First 3 seconds spoken. No product name in the first 5 words unless it is the pattern interrupt."
Required mix: 3 problem-callout · 3 curiosity · 2 social-proof (only if proof pack has numbers) · 2 pattern-interrupt visual directions · 2 myth-bust
Output schema: CSV rows: id | archetype | hook_line | visual_direction_0_3s | target_funnel (cold/warm)
Visual direction field: One shot instruction: face, product enter, overlay text ≤5 words, or none
Forbidden in hooks: Questions that dump the category ('looking for a better [x]?') · 'wait for it' · 'link in bio' · invented stats
Selection rule after output: Pick 3 hooks from different archetypes · discard any line that needs a fact not in the pack
Follow-up if weak: "Rewrite hooks 1–12 using these exact review phrases: [paste]. Do not paraphrase them into marketing English."

30s timed-script prompt

User prompt stem: "Expand hook #[id] into a 30s talking-head UGC script. Skeleton: Hook 0–3 / Problem 3–8 / Agitate or proof 8–16 / Demo 16–24 / CTA 24–30."
Word budget: 75–90 spoken words total · mark [OVERLAY: …] separately · 1 spoken CTA max
Output schema: Rows: t_start-t_end | spoken | overlay | on_camera (face/product/insert)
Proof rule: Every numeric claim must cite the pack field it came from in a trailing #source column · if none, delete the number
Mute-first rule: Overlays must carry the claim if sound is off · max 5 words per card · change card every 3–4s
CTA lock: Spoken CTA must match the live offer field · if offer is empty, CTA = "link below" only, no discount
Length fail: If spoken >95 words, regenerate with "cut 20% , keep hook and CTA" — do not speed-read

3×2 variation prompt (one variable per pass)

Grid definition: 3 hooks × 2 deliveries (calm vs punchy, or 2 identities) = 6 scripts · body locked
User prompt stem: "Keep seconds 8–30 identical. Write 3 new 0–8s opens from archetypes [A,B,C]. Then duplicate each in [delivery_1] and [delivery_2] tone notes."
Output schema: filename_stub | hook_id | delivery | 0_8s_spoken | 0_8s_visual | rest_unchanged=YES
Illegal mix: Do not change offer, proof number, or CTA inside a hook test · that contaminates the cell
Offer test (separate pass): Lock hook+body · swap only CTA overlay and spoken close · 2 offer cells max
Identity test (separate pass): Same script · label PEER vs FOUNDER · never both in one ad set
Batch cap: 6 files per prompt · if you need 12, run the prompt twice with new hook ids, do not ask for 12 at once

Claim-safe QA rewrite prompt

User prompt stem: "Audit this script against the banned-claim list and Meta/TikTok personal-attribute patterns. Rewrite only failing lines. Return a diff."
Fail if — personal attributes: "you look tired / your hormones / I know you're in debt" · rewrite as speaker's own past, not viewer's trait
Fail if — unsubstantiated result: Any number, before/after, or 'my doctor' not in the pack · delete or replace with mechanism
Fail if — guarantee mismatch: Script mentions refund/trial/"results in 7 days" but offer pack does not · strip the line
Output schema: original | issue_code | rewrite | still_needs_human (YES/NO)
Pass stamp: All rows still_needs_human=NO · word count still 75–90 · hook still first sentence
Human gate: Health, finance, kids, performance claims · still_needs_human=YES always · model may not clear these

Locked

Unlock the full cheatsheet

Enter your email and we'll instantly unlock the complete cheatsheet — plus send you fresh ad playbooks every month.

No spam. Unsubscribe anytime.

Ready to put this into practice?

Create your first AI UGC video ad in minutes — no filming, no actors, no editing.

Try Klip Kanvas free