The Death of the Essay-Prompt

The prompting trap

If you look at most AI tutorials on the internet, the advice is always the same: write a better prompt. More context, more detail, more paragraphs. Specify your audience, define your tone, list your brand guidelines. Write an essay just to get one asset out.

That’s not automation. That’s just a different kind of manual labour.

This week both pipelines hit the same milestone: stop asking the user to brief the AI, and start building the brief into the system itself. The first step was locking down the brand intelligence schema based on everything testing had surfaced. Colours, fonts, logos, exclusion zones, tone, image style rules. Defined once, structured properly, reusable everywhere.

Narrowing the tube on the web

The gap between a generic output and a useful one came down almost entirely to the brief. The first version was too broad — industry listed as “cloud computing,” audience listed as “SMBs and enterprise.” Those labels describe a market. They don’t give the model enough to make a real design decision with.

Once the inputs got more specific, the output improved noticeably. But that surfaced the real problem: clients can’t be expected to write detailed AI prompts. Most people don’t think in terms of prompt engineering, and they shouldn’t have to.

So the tool was redesigned around that constraint. During onboarding, users provide information through simple inputs — a brand brief, tone preferences, audience selections. Each choice feeds into a larger prompt that gets assembled automatically in the background. The user makes quick, intuitive decisions. The platform translates those into the detailed instructions the model actually needs.

The main takeaway was that the quality of AI-generated output is largely determined before the model is even used. Designing a process that gathers the right information without increasing the workload for the user — that became the real challenge of the project.

Engineering the straightjacket for images

On the image side, the focus this week was on how the tool actually gets used — and what that process looks like when most of the work is already done before the user types anything.

A fully editable layout generation request constructs a prompt of around 6,000 words behind the scenes. A baked image sits closer to 500. The user writes a few words, maybe a few sentences. The system fills in the rest — brand context, layout logic, visual style rules, tone — all assembled automatically into something the model can actually work with.

The way that gets built is through layers. Built-in styles hold JSON information about layout configurations and visual combinations that have been tested and work well. Built-in prompt extensions handle the stylistic direction — mood, composition, treatment — so the user can click rather than describe. The brief constructs itself from selections, not sentences.

Templates work the same way. Each one stores information about what it’s used for and what all its components are — headline, subheading, image area, CTA, spacing logic. So when a new version needs to be generated, the system already knows what’s there. A human or an automated reprompt only needs to specify what’s changing. The AI knows what to update and what to leave alone. No full rebriefing, no rebuilding from scratch.

And because the editable layout approach keeps elements live on the canvas rather than baking them in, tweaks happen without spending tokens. The prompt structure is also built to be efficient, so when a generation does run, it’s not burning through context unnecessarily.

Systems over text boxes

The shift across both projects is the same: the thinking moves out of the prompt and into the infrastructure.

The brand schema means the model isn’t guessing. The built-in styles mean the user isn’t writing essays. The template memory means nothing gets rebuilt from scratch when it doesn’t need to be.

What’s left for the user is the part that actually requires a human — what this creative is for, what it needs to say, what’s different about this one. Everything else the system already knows.