The AI workspace for marketers: briefs, drafts, and image generation in one tool

Why most AI marketing copy is unusable

It is not the model. It is that "write a landing page for our product" gives the model nothing that distinguishes your product from a thousand others, so it returns the average of all of them. That average has a recognisable texture: inflated adjectives, benefits with no mechanism, and a headline that would fit any competitor if you swapped the logo.

Everything on this page is downstream of one fix. Give the model three things before you ask for copy: how you sound, what is true, and what is forbidden. Everything else is technique.

Task	Model	Why
Long form copy, brand voice, email	Claude	Warmest register, least prone to hype, holds a voice across a long piece
Briefs, variant sets, structured output	GPT	Follows format constraints, generates n clean variants without drifting
Competitor and category research	Gemini	Recent web material with sources you can verify
Campaign imagery and social assets	Flux, Stable Diffusion	Included in Whizi, so no second image subscription
Choosing between two drafts	Both, side by side	The cheapest quality lift available

Build the brand voice file once

This is a 30 minute task that improves every prompt you write afterwards. Create one document and paste it into the top of any copy prompt. It should contain:

• Three to five paragraphs of your actual best copy, unedited. Real examples beat any description of tone.
• A banned list: the specific words and constructions you never use. Most teams have them. "Revolutionary", "seamless", "game-changing", "unlock", "in today's fast-paced world".
• Your proof points, written as facts with numbers. "Cuts review time from three days to one for two pilot teams" is a proof point. "Saves time" is not.
• Your claims boundary: what legal will not let you say, and how you phrase the near version instead.
• Two or three sentences on the reader, specifically what they already believe and what they are sceptical of.

Prompt: extract the voice from your own archive

Read these examples of our writing. Describe the voice in operational terms a writer could follow: sentence length pattern, level of formality, how we open, how we handle claims, what we never do. Then list ten words or phrases that appear repeatedly and ten that never appear. Examples: [paste 3 to 5 pieces].

Save the output. That description plus the raw examples is the most valuable prompt asset a marketing team can own, and it is the difference between AI copy that needs a rewrite and AI copy that needs an edit.

Campaign brief to first draft

Prompt: the brief

Draft a campaign brief. Product: [what it does, in plain language]. Audience: [specific segment, not "businesses"]. The belief we need to change: [what they currently think]. Evidence we can use: [proof points]. Channels: [list]. Constraint: [budget, timing, legal]. Return: the single message, three supporting angles, the objection each angle handles, what success looks like as a measurable outcome, and what we are deliberately not saying.

The last field is the useful one. Naming what the campaign will not claim keeps the brief from becoming a list of everything good about the product, which is the standard way a campaign loses its edge before a word of copy is written.

Prompt: messaging angles before polish

Generate five distinct positioning angles for this campaign. Distinct means they rest on different reasons to believe, not different wordings of the same reason. For each: the angle in one sentence, who it lands hardest with, the proof required to support it, and the strongest objection to it. Do not write finished copy yet. Brief: [paste].

Prompt: ad variants

Write [n] ad variants for [channel]. Character limits: [exact limits]. Voice: [paste brand voice file]. Angle: [chosen angle]. Each variant must lead with a different mechanism, not a different adjective. No claim beyond the proof points listed. Return as a table: variant, hook type, the proof point it uses, and the objection it pre-empts.

Asking for the hook type and objection in the table is what makes the set testable. Twelve variants that vary only in phrasing teach you nothing when the test finishes. Twelve variants built on different mechanisms teach you which mechanism works.

Landing pages, email, and the edit that actually matters

Prompt: landing page

Write a landing page for [product] targeting [segment]. Structure: hero (headline, subhead, one CTA), problem section in the reader's own language, how it works in three steps, proof, objection handling, and final CTA. Rules: the headline must state a specific outcome, every benefit must name its mechanism, no statistic that is not in the proof points, no more than one exclamation of enthusiasm in the whole page. Voice: [paste voice file]. Proof points: [paste].

Prompt: lifecycle email

Write a [type] email for [segment] whose last action was [behaviour]. Goal: [single action]. Tone: [from voice file]. Constraints: under [n] words, one call to action, subject line under 45 characters, no fake urgency, no "just checking in". Return three subject line options with the trade-off each one makes.

Prompt: the criticism pass, before any rewrite

Before rewriting, identify the three weakest things about this draft: a claim with no proof, a sentence that would fit any competitor, a benefit with no mechanism, or a tone mismatch against the voice file. Quote each one. Do not rewrite yet.

That criticism prompt is the highest leverage habit in this whole workflow. Models will polish forever if you let them, producing smoother copy that says nothing new. Forcing an evaluation step before the rewrite gets you a different draft rather than a shinier one.

Then compare. Run the same brief through Claude and GPT side by side and keep the better opening, the better proof section, and the better close. That is two minutes of work and it reliably beats iterating on a single draft.

Campaign imagery without a second subscription

Image generation is included in Whizi on Pro and above, which for most marketing teams removes the Midjourney line item. The models that matter are Flux for photographic and product-adjacent work, and Stable Diffusion variants for stylised or illustrative treatments.

The prompt pattern that produces usable assets rather than pretty noise:

[subject], [action or state], [setting], [lighting], [composition and framing], [lens or medium], [colour treatment], [mood]. Negative: [what must not appear].

Worked example: A mid-thirties operations manager reviewing a tablet in a warehouse mezzanine, morning light through high windows, shot from slightly below at eye level, 35mm, muted blues and warm greys, calm and competent. Negative: stock-photo smiling, cluttered background, visible logos, text.

Practical notes from real campaign work:

• Generate at the aspect ratio you actually need. Cropping a square into a banner wastes the composition you asked for.
• Iterate one variable at a time. Change the lighting or the framing, not both, or you will not know what fixed it.
• Keep text out of generated images and add it in your design tool. Rendered text is still the weakest area across every image model.
• Check your usage rights and your brand rules before anything generated goes into a paid placement. Faces, recognisable locations, and anything resembling a real person deserve particular caution.
• Save the prompts that worked. A prompt library of eight reliable house styles is worth more than any single image.

Competitive and category monitoring on a schedule

This is the task most teams intend to do monthly and actually do never. It takes fifteen minutes with a web connected model and a saved prompt.

Prompt: the monthly scan

Research what changed for [competitor list] in the last [period]. Cover: messaging and homepage positioning changes, pricing or packaging changes, new product or feature announcements with dates, funding or acquisition news, and any shift in the segments they appear to be targeting. Cite every claim with a URL and a date. Where you cannot verify something, say so rather than inferring.

Prompt: the positioning map

Using the scan above, place each competitor on two axes that genuinely differentiate this category (propose the axes and justify them). Then state where we sit, what space is currently uncontested, and what would have to be true for that space to be worth owning.

For a deeper treatment of sourcing and verification, see the guide on AI for market research.

Does the AI copy actually perform?

Worth being honest about. AI accelerates production, and production was rarely the bottleneck. What it changes is how many genuinely different angles you can afford to put into a test, which is where the performance gain comes from. Twelve variants of the same idea will not beat your control. Four variants resting on four different mechanisms might.

Two habits keep this from becoming volume for its own sake. First, tag every variant with the mechanism it uses, so the test result tells you something transferable. Second, feed the winner back into the voice file as a new example, so the next generation starts from what worked rather than from the average of the internet.