A campaign is a sequence of jobs
The brief is a blank page, the channel is not chosen, and someone has already asked for "ten hooks by noon." Marketing tools that write hooks on an empty brief produce hooks you cannot defend. The useful AI workflow follows the campaign, not the model's appetite for copy. Research, then a claim you are willing to make, then drafts, then variants, then a human who can be blamed in a good way if the ad is wrong.
Nothing here is a promise about conversion rates. This site has not gone and measured your funnel. If a vendor tells you a model lifts revenue by a number they cannot show you the study for, ask for the study. If they cannot produce it, ignore the number.
Write the claim before you write the ad
One sentence, in a doc, in your words: what the product does for whom, and what you will not imply. "We ship a weekly box of coffee beans to offices in the city. We do not claim health effects. We do not claim the beans are rare." That sentence is the source. Every draft has to be able to point back to it.
If you do not know the claim yet, you are still in research. Use a cited tool or the primary pages, as the Perplexity review describes, and talk to the person who owns the product. A chat assistant is a bad place to discover your own offer. It will average the category and hand you a claim that sounds like every competitor, including the ones you dislike.
Jasper and similar marketing suites exist because teams got tired of pasting brand rules into a general chat. They add workflows, templates, and a place to store a voice. They are worth the setup when several people publish and the voice document is real. They are overhead when you are one person and the voice document is a sentence. A general assistant plus a pinned note may be the whole stack. Read Jasper's current plan page if you are comparing costs. Do not buy a suite to avoid writing the sentence.
Research that cannot invent the customer
Collect the words customers already use. Support tickets, reviews you have the right to read, sales call notes. Paste a sample into the assistant and ask for themes, with the instruction to quote the lines it groups. You are looking for phrasing, not for a fictional persona with a name, a city, and a sad commute. Invented personas are how ads start talking to nobody.
Competitor pages are public. You may summarize what they claim. You may not lift their sentences. Ask the model to list claims, then you decide which claims you refuse to make even if they are common. "Fastest" and "only" are claims. If you cannot prove them, they do not go in the draft, no matter how often the model offers them.
Drafts, variants, and the difference
Once the claim and the phrases exist, a writing model is in its lane. Give it the claim, two samples of your real ads that performed in the boring sense that you were willing to run them, the channel, and the length. Ask for a small set of variants, not fifty. Fifty is a way to avoid choosing.
Read variants for one thing: do they say something the claim sentence does not allow. Cut those first. Then read them aloud. Marketing copy that only works on the screen often hides a stumble.
The prompt guide is the mechanics of that brief. The writing tools guide is the longer version if the asset is an article rather than an ad. Channel changes the shape. An email and a paid search line are not edits of each other. Ask for one shape per turn.
Claude is a reasonable place for a long landing page that has to follow a source doc. ChatGPT is a reasonable place when the same claim must leave as a page, a short ad, and a caption, and you want those as separate, labeled blocks in one thread. Pick one default so the team is not maintaining voice notes in four products.
Pictures and the claim have to match
Generate images after the headline is stable. A picture of a doctor, a lab, or a "result" is a claim even with no words. The image comparison covers where the picture is made and why the license belongs in the project folder. For ads, the platform's own rules on synthetic media sit on top of the model license. Read the platform page the week you publish. Policies move.
Do not generate a customer's face. Do not generate a fake before-and-after. Do not invent a testimonial and then make a portrait to match it. Those are not gray areas in a serious brand. They are ways to lose the account and, depending on where you advertise, to break a rule that carries a penalty.
Review is a role, not a vibe
Someone who did not write the draft should read it against the claim sentence and against the offer's real terms: price, geography, what is included, what the refund path is. The model does not know the refund path unless you pasted it. If you did not paste it, the draft should not mention refunds.
Legal and compliance review, if your category needs it, happens on the human draft. Do not ask a chatbot whether an ad is "compliant" and treat the answer as clearance. That question feels efficient and it is not advice. Send the ad to the person who is paid to answer it, with the claim sentence attached so they are not reviewing a vibe.
Track what you changed. If every draft loses the same sentence in review, add that sentence to the do-not-say list in the brief. The list is how the tool gets more useful without a new subscription.
Measurement stays outside the chat
The assistant can help you write a hypothesis: "We think the headline that names the city will be clearer than the headline that names the roast." It cannot tell you the hypothesis won. Your ad platform and your analytics can, if they are installed correctly and you give the test enough time. Ask the model to help you describe the test in plain language for a colleague. Do the arithmetic in the sheet, or check any arithmetic it offers against the sheet.
Bring results back as words customers used, not as a victory lap for the model. The next brief should quote the reviews and the claim, not "the AI is working."
A stack a small team can actually share
One shared doc for the claim, the do-not-say list, and three real samples. One writing assistant. One place for images, with prompts saved beside files. One cited-search habit for anything that sounds like a fact about the world. A calendar reminder to reread platform policies when you start a new channel.
Add a marketing suite when the shared doc is being ignored because people will not leave their tool. Do not add it as a substitute for the doc. The directory of tools can suggest names. Your claim sentence decides which names are relevant.
If you are a one-person business, the small business guide starts from the inbox rather than the campaign. Same rule: the source of truth is yours, and the draft is a draft.
Noon, with the hooks
The ten hooks exist, and each one can be read next to the claim sentence without blushing. The ones that cannot, do not go to the media buyer. You still have to choose, which is the job the model was never going to take. Choosing is the marketing. The rest is typing you were happy to speed up.