How to Use Claude for Brand-Safe Ad Copy and Long UGC Scripts

Write a pinned system prompt that states who the speaker is, who they are not, what they may claim, what they must never claim, and how they talk. Put the banned list, tone rules, and audience in the same block. Claude follows a dense, explicit constitution better than a mood board. If the safety rules live in your head, they will not live in the draft.

Include five sections, labeled: Speaker, Audience, Allowed claims, Forbidden claims, Voice. Speaker is a customer or creator, not the brand. Audience is one buyer, not 'everyone.' Allowed claims are copied from your one-pager, not inferred from the website. Forbidden claims cover health, income, competitor defamation, 'guaranteed,' and anything in review last quarter. Voice is concrete: contractions on, no tricolons, no 'unlock your potential,' first person for UGC, second person sparingly. Paste this into a Claude Project so every thread inherits it.

Add a refusal rule in plain language: 'If a request would require a claim not on the allowed list, do not paraphrase a loophole. Say what you cannot say and offer an experience-line substitute.' Claude is unusually willing to honor that if you say it once, at the top, and keep it there. Do not bury safety under twenty examples of winning ads. Examples teach style; the constitution teaches boundaries. You want both, with boundaries first, or the model will imitate a bold competitor ad you cannot legally run.

Paste a table with three columns: claim as spoken, evidence, and the hedge you will use if evidence is thin. Then paste five to ten customer quotes. Do not dump the product page. Claude will treat every sentence on a PDP as fair game unless you tell it the table is the only source of proof. Websites over-claim in ways you would never put in a paid script. Your ads should not inherit that extra certainty.

Each row should be short enough to say in one breath. 'Ships in two days if ordered before 2pm' with a link to the shipping policy is a usable row. 'Transforms your skin' with a blog post is not. Where evidence is a survey percentage, write the exact figure and the sample caveat in the hedge column so Claude cannot round it up. Where evidence is anecdotal, the spoken line must stay first-person and time-bounded ('for me, about a week'). That mapping is the whole point of using Claude for brand-safe copy: you are asking it to stay inside rows, not to invent a better row.

If you sell in multiple markets, add a column for where the claim is cleared. A supplement line that is fine in one country can be a health-claim problem in another. Claude will happily reuse a US script for a UK landing page unless the table says not to. Keep the table in the Project files, not only in one chat, and tell the model to cite the row it used in a hidden note under each draft. Those notes make the legal review in a later step a lookup instead of a hunt through a long thread that nobody else can search.

Give Claude the four-beat UGC spine, then ask for a 45-to-90-second script (about 110 to 220 spoken words) that can carry demo detail, an objection, and a specific CTA without rushing. This is where Claude earns its keep: long context, fewer dropped constraints, and a better chance the middle does not collapse into bullet-point benefits. Keep hook exploration in a tighter, separate pass.

Long UGC is a different object from a 30-second hook test. You need room for a real objection ('I thought this was just another bottle'), a visible demo, and a recap that does not repeat the hook. Prompt for labeled beats so you can cut one without rewriting all of them. Ask Claude to mark optional sentences in brackets so a 90-second cut can become a 45-second cut without a new generation. That bracket trick is more useful than asking for 'a short version' after the fact, which usually produces a different script instead of a trim.

For landing-page or email copy, change the speaker rules in a child prompt: second person is allowed, first-person UGC is not required, but the claims table still governs. Ask for a headline, a subhead, and a body that does not introduce a proof point absent from the table. Claude will otherwise 'help' by pulling a study-shaped sentence from general knowledge. Your system prompt's refusal rule should catch that; if it does not, tighten the rule and regenerate. Brand-safe means boring on proof and specific on experience, not the reverse.

In the same thread, ask Claude to list every factual statement, map each to a row in the table, and rewrite orphans as experience lines or delete them. Do this before you fall in love with a turn of phrase. Fluency is how unsafe claims sneak in, because the sentence sounds finished. The audit is the product you paid for. The script is whatever is left after the orphans are gone.

Specify the output shape: a two-column list of Spoken line / Source row or 'no row — rewritten.' Then have it emit a clean script with the orphans already fixed. If you only say 'make sure this is on-brand,' Claude will compliment the draft and change almost nothing. Brand-safe work is clerical. Treat it like clerical work. You should be able to hand the audit table to legal without rewriting it. If you cannot, the prompt was still a vibe, and the draft is still unreviewed copy wearing a confident tone.

Watch for implied claims, not just explicit ones. 'My dermatologist was shocked' implies a medical endorsement you probably do not have. 'This is the only one that worked' is a superlative plus a competitor comparison. Claude will flag these more reliably if you give three examples of implied claims you have had to kill before. Models learn your risk pattern from samples. One round of audit is not paranoia; it is cheaper than a rejected ad set or a landing page that has to be recut because the voiceover promised a result the PDP does not.

Once the long body is clean, ask Claude for three offer variants and three objection variants with the claims table still in force. Change one variable per run: offer, objection, or proof row. Do not ask for 'fresh angles' as a blob. Fresh angles are how a safe script grows a new unsubstantiated miracle in draft four, usually in a sentence you will miss because the rest of the page already passed.

Name the variable in the prompt: 'Keep beats 1, 2, and 4. In beat 3, swap the demo from texture to time-to-result using only row 4.' That is a testable variant. 'Make it punchier' is not. Punchier is usually more absolute, which is the opposite of brand-safe. If you need a harder hook, write the hook separately and run it through the same audit. Do not let a punchy opening smuggle a new claim into a body you already cleared and numbered. Keep beats 1, 2, and 4 frozen while that one beat moves, so reporting can name the variable without a detective.

For ad copy sets — RSA headlines, Demand Gen headlines, email subjects — instruct Claude to stay inside character limits and to reuse proof language rather than invent synonyms that change the meaning. 'Clinically tested' is not a synonym of 'clinically proven.' 'Up to' is not optional decoration. Paste the character limits and ask for a discarded pile: lines it considered and rejected for safety, with a one-line reason. That discarded pile is training data for your next system-prompt revision, and it is more useful than another round of praise for the keepers.

Do not send Claude's entire afternoon to legal. Send the locked long script, the three variants you will actually shoot, and the audit table. Humans catch tone problems models rate as fine: a joke that punches down, a health-adjacent wink, a competitor nickname. Claude reduces volume so reviewers can read. It does not hold the risk if a line ships.

Give reviewers a simple rubric: claim true, claim approved for this market, speaker allowed to say it, tone on-brand, nothing implied that we would not print. If legal only has time for one pass, they should see the spoken lines and the on-screen captions together. Split reviews are how a safe voiceover gets an unsafe lower-third. Keep the Claude thread link or exported audit with the file so a later question is answerable without reconstructing the chat from memory a month after the campaign launched.

If a reviewer changes a line, paste the change back into Claude and ask it to propagate the same constraint through the other variants. That is a better use of the model than asking it to 'make it legally safer' in the abstract, which produces hedges that weaken the ad without matching what counsel actually wanted. The human sentence is the new source of truth. Claude's job is to spread it, not to negotiate it back toward the original flourish. Counsel wrote a sentence; copy that sentence into every variant that shared the risk.

Format the long script with pause marks, attach the approved captions, and note which claims row each demo beat leans on. Then generate or film. If production improvises, they must stay inside the table. Brand-safe copy dies in the gap between a cleared doc and a 'small tweak' on set that nobody runs back through the audit, and the cleared version dies in the export.

Klip Kanvas can render the cleared long-form UGC with a consistent avatar and captions you paste from the reviewed file, which is the point of doing the claims work in Claude first: production should not be a second writer. Watch the full cut once for restored superlatives in captions and for pacing that makes a hedged line sound like a guarantee. If you need a 30-second child cut, trim the bracketed sentences you marked in step three rather than asking any model for a brand-new short script that will need another legal loop.