AI makes it possible to see a designed page almost immediately. That creates an obvious temptation: why plan first when we can just generate the page and react to it?
We tested whether a planning step actually changes the quality of the work, rather than simply making the process feel more deliberate. We gave the same AI the same fictional company brief and asked it to create the same homepage two different ways for Meridian Peak Capital.
The setup
Design immediately
The model received the Meridian Peak Capital brief and generated the homepage with no separate planning stage.
Plan before design
Before designing anything, the model determined page objective, visitor state, positioning, message hierarchy, audience priorities, proof requirements, factual constraints, objections, narrative sequence, and section roles. It then designed the homepage using that plan.
The planned homepage was more coherent, more disciplined with facts, and stronger at preserving the relationship between a claim and the information supporting it. But the design-first version produced some of the stronger individual creative decisions. That tension is the central story of this article.
What the planning step decided
Results: design immediately
The design-first version produced the stronger individual headline:
“Your strategy doesn't need to change. Your decisions get harder.”
It was sharper and more distinctive than the planned version's safer opening.
It also elevated Meridian Peak's six operating areas into a dedicated “Where we get involved” section. Instead of becoming a generic SaaS card grid, the section was executed as a restrained editorial list that made the firm's operating role concrete and scannable.
- Go-to-market strategy
- Leadership development
- Pricing
- Market expansion
- Operating infrastructure
- Strategic decision-making
The evaluators considered that a meaningful strength.
But the design-first version made weaker strategic decisions elsewhere. It placed the philosophy before the investment criteria, meaning a time-constrained founder encountered an abstract belief before learning whether Meridian Peak was relevant.
More importantly, it dropped the strongest credibility fact supplied in the brief: Meridian Peak was founded by investors and operators. The page asserted that strong companies need experienced partners but omitted the strongest available factual reason to believe Meridian Peak could provide that experience.
The design-first page also had production issues including mobile navigation disappearing and visible contact placeholders. These affect launch readiness but should not be attributed to the absence of planning. They are execution and QA failures.
Results: plan before design
The planned version was more disciplined. Its sequence was:
It preserved the founder-and-operator fact and placed that fact directly beside the philosophy it supported. The source-aware evaluator identified this as the clearest planning-related improvement.
Planning also made missing proof explicit before design. Because portfolio companies, named team members, fund statistics, testimonials, and press had not been supplied, the plan excluded those elements instead of allowing plausible-looking evidence to be invented.
But planning did not improve everything. The plan predicted that giving the six operating areas their own section might create a generic SaaS-style card grid, so it compressed those ideas into prose. The design-first output demonstrated that this predicted failure was not inevitable: it created a restrained editorial list that evaluators preferred for specificity and scannability.
Results: what planning actually changed
“Planning helped most where it constrained facts, proof, objectives, and argument structure. It helped least when it tried to pre-decide creative form before the alternative had actually been designed.”
A plan can contain bad decisions too
The plan was also generated by AI. It can therefore contain strong judgments, weak judgments, assumptions, overcorrections, and reasonable-sounding predictions that have not been tested.
- “Put criteria before philosophy” was a useful strategic judgment.
- “Keep the credibility fact beside the philosophy it supports” was useful.
- “Do not elevate these six operating areas because the result may become a generic grid” was a creative hypothesis, not a fact.
The actual design demonstrated that the six items could be elevated without producing the predicted failure.
Constraints vs. hypotheses
What should strongly constrain generation
Known company facts, factual unknowns, audience requirements, explicit objectives, proof relationships, and anti-fabrication rules.
What deserves strong consideration
Narrative sequence, section roles, message prominence, and where objections should be resolved. These remain open to evaluation.
What should remain challengeable
Prose versus list, whether something gets its own section, composition choices, and predicted visual failure modes. These are inputs to exploration, not binding instructions.
The evolved workflow
- Context defines the company and problem.
- Planning makes communication decisions explicit.
- Plan evaluation tests those decisions before they become constraints.
- Generation explores the creative execution.
- Output evaluation determines whether the work actually solved the problem.
- Refinement applies what was learned.
“A plan should not become truth simply because the AI wrote it before it designed.”
Generation benefits from planning.
Planning benefits from judgment.
Methodology limitations
This was one fictional company, one homepage task, one model/version, and one pair of generated outputs. A blind evaluation was followed by a source-aware evaluation using the original brief and planning artifact. The findings describe what happened in this test and should not be generalized across all models, companies, or creative tasks without repeated testing.