The Prompt Asks Before It Builds
The most common failure mode with AI-generated marketing content is not that it writes badly. It is that it writes confidently about things it invented.
We ran into this problem while building module 3 of the Freedom Ag training portal. The module's job was to give the client five prompts they could use in any Claude session: a workflow cost audit, a landing page builder, a five-email sequence writer, a short video script, and a grower-acquisition thinking partner.
Each one had the same constraint: we could not let the prompt build anything before it understood the business. A landing page built on guessed claims is not just generic. Made-up testimonials or fabricated star ratings on a real business page can get the owner in actual trouble.
So every prompt in the pack opens with the same hard rule, stated plainly: never generate from nothing. Interview first.
What "interview first" looks like in practice
The bottleneck finder asks six questions about the owner's week before it runs any math. The landing page builder works through eight interview steps before writing a single line of HTML. The email writer asks for two real emails the owner has sent, to learn their actual voice, before drafting anything.
The interview is not just courtesy. It is the only thing standing between the output and fabrication.
Each prompt is also explicit about what fabrication looks like, so the model knows what to stop itself from doing: no invented testimonials, no made-up review counts, no star ratings the owner did not confirm. If proof does not exist in what the owner said, it does not appear in the output.
The grower-acquisition prompt
The most constrained prompt in the pack is the grower-acquisition thinking partner. Ag retail is a relationship business. Growers talk to each other. One pushy move costs a reputation that took years to build.
The prompt's job is to generate marketing ideas that would survive what we called the coffee-shop test: if two growers mentioned an outreach to each other, it should make the business look helpful rather than hungry. That one framing eliminated entire categories of common marketing tactics without us having to enumerate them.
We structured it as three interview rounds before any ideas are generated. Round one covers who the best growers are and why they buy from this business instead of the co-op. Round two covers where new customers have actually come from in the last two years, and what the owner tried that felt gross. Round three covers what the business knows that growers find genuinely useful: agronomy knowledge, local trial data, service stories.
Only after all three rounds does it produce five ideas. Each idea has four required parts: the concept itself, why it is not salesy, why it might fail, and what the first step and 90-day scoreboard look like.
The last section is called "The One I Would Kill." That is the one common ag-marketing move the prompt is deliberately not recommending, and why it would backfire with this specific client's growers. It exists because recommending what not to do is often more useful than listing what to try.
Prompts as deliverables
These prompts live on the training portal as copyable text. The client pastes any of them into a fresh Claude session and runs it. They are not embedded in an app or locked to a workflow. They work wherever Claude runs.
That choice was deliberate. The training portal has videos, tenant auth, and a private blob delivery system. But the prompts themselves are just text. Making them copyable means the client owns them in the most literal sense. They can modify them, share them with staff, and use them without anything we built.
The portal is how they learn to use the prompts. The prompts are the thing they take away.