The Blank Prompt Problem: What to Do When Claude's Ideas All Sound the Same

You open a fresh Claude session, type "give me ideas for a landing page for my product," and get back a list so generic it could belong to anyone. Hero section. Social proof. Clear CTA. Value proposition above the fold. Ten ideas, all interchangeable, none of them yours.

This isn't a model problem. It's a context problem — and it's the single most common failure point in the whole AI-assisted build pipeline, because it happens at step one. If ideation returns nothing usable, everything downstream inherits the blandness. The page gets built, the blog gets scaffolded, the automation runs — all efficiently producing something indistinguishable from a thousand other sites.

Here's how to fix the input side so the output side has a chance.

Why "give me ideas" fails

An LLM with no context does the only rational thing: it averages. It reaches for the most statistically common answer to your question, which by definition is what everyone else gets. The more open the prompt, the more the response regresses to the mean.

Compare two prompts:

"Give me ideas for a landing page."

versus

"I'm building a landing page for a scheduling tool aimed at
solo barbers. The competing tools all market to salons and
spas. My differentiator is one-tap rebooking via SMS. Audience
finds me through TikTok. Give me five angle-of-attack options
for the hero headline, each targeting a different pain."

The second prompt isn't longer because more words are magic. It's longer because every clause removes one direction the model could have drifted in. Constraints are what generate ideas, not freedom.

The three context layers that change everything

When a brainstorm comes back flat, the fix is almost always one of three missing layers.

1. The constraint layer: what it is NOT

Before you ask for ideas, list the exclusions. "Not a SaaS dashboard aesthetic. Not blue. No 'Streamline your workflow' phrasing. No free trial CTA — we qualify leads on a call." Models are dramatically better at avoiding things than you'd expect, and exclusions are cheap to write. A short "don't" list will kill half the generic output on its own.

2. The proof layer: what actually happened

Generic ideas are generic because they're unmoored from anything true about your business. Feed in the raw material — a customer email, a sales call transcript, the review that made you laugh, the support ticket that keeps recurring. Then prompt like this:

"Here are three real customer quotes. Generate hero headline options that each make one of these quotes the claim, and tell me which quote each one leans on."

The output stops being "Transform your productivity" and starts being something only your product could honestly say. This is also your fact-checking anchor: if the model invents a claim, you can see it immediately because it doesn't trace back to anything you gave it.

3. The adversarial layer: make it argue with itself

Flat ideation often comes from asking one question once. Instead, run the session in rounds:

1. Round one: "Give me five angles for this page. Flag which one you think is weakest and why."
2. Round two: "You picked X as weakest. Steelman it anyway — who would it win with?"
3. Round three: "Now write the version of angle Y that would embarrass angle X."

This works because a single pass optimizes for plausibility, but a critique pass forces the model to surface tradeoffs it smoothed over the first time. The same trick works for blog ideation — ask for ten post ideas, then ask which three a skeptical reader would call filler and why, then regenerate those three.

Prompting for structure, not sentences

One trap worth naming: don't ask the model to write the final headline in the ideation phase. Ask it to write the strategy of the headline. "The hero leads with the pain of no-shows and positions SMS rebooking as the fix, in under 12 words." Then you write, or let the model write, the actual line from that spec. Ideas drafted as finished copy get reviewed as finished copy — and finished copy triggers the editor brain that kills everything. Ideas drafted as strategy get improved.

The same applies to site structure. If you're starting from a campaign brief, the pre-build step of turning that brief into testable wireframes before any code exists is exactly this principle applied to layout — we walked through that workflow in "Before Claude Writes a Line of Code: Wireframing Your Landing Page With LLMs Using Only Your Campaign Brief". Ideation and wireframing are the same discipline at different altitudes: constrain first, generate second, critique third.

When the model is still stuck: change the input, not the model

If you've layered all three context types and the output is still beige, stop re-prompting and change the input format:

• Paste, don't describe. Actual transcripts beat your summary of them. Summaries strip the specificity the model needs.
• Swap roles. "You are a skeptical prospect who has seen 40 landing pages this week. React to my current homepage text." Reactions generate better ideas than requests.
• Force a format. "Give me every idea as a one-sentence pitch with the target customer and the risk attached." The risk field alone will surface ideas the freeform version never produced.
• Ask for the boring version. "What would the most obvious, safe version of this page be? Now list everything that makes it obvious and safe." That list is your map of the differentiation space — the obvious version is what everyone else is already shipping, and your job is to be measurably not that.

Switching models — Claude to GPT or back — sometimes helps marginally, but it's the last lever, not the first. A better-context session on the same model beats a better model on the same empty prompt every time.

The takeaway

Bad ideation output isn't a reason to distrust LLMs in your build pipeline. It's a sign you're using them as an oracle instead of a collaborator. Oracles need nothing from you. Collaborators need the constraints, the proof, and the pushback that make a real brainstorm work — the same things that make a good human hire good.

Spend twenty minutes building the context once, and you can reuse it across every page and post you generate after. Which is the whole point: the pipeline only pays off if the ideas at the top of it are worth automating the production of.