Behind the Scenes: How I Test Prompts for formspace.design

As the founder of formspace.design, I am personally involved in developing many aspects of the product, including how our artificial intelligence understands your requests and generates landscape concepts. This isn't just about writing a few lines of text; it's a deep, iterative process aimed at creating truly useful and implementable solutions for every plot. My goal is to ensure that the AI in formspace.design is not just an image generator, but a reliable assistant that offers well-thought-out and functional designs.

Why Prompt Quality is Critical for formspace.design

My many years of experience in landscape design have taught me that "beautiful" is only part of the equation. A project must be functional, economically viable, and most importantly, implementable. When I first started working with AI, I quickly realized that the quality of the output directly depends on the quality of the input – that is, the prompt. If a prompt is imprecise, the AI can generate anything from fantastic but impossible-to-realize solutions to completely impractical options. It's important to me that formspace.design offers solutions that I myself would confidently implement on a client's site. Therefore, every prompt that forms the basis of our AI modules undergoes rigorous testing.

My Step-by-Step Process for Developing and Testing Prompts

Developing effective prompts is not a one-time task but a continuous cycle of improvements. Here's how I approach this process:

1. Clearly Defining the Goal

Before writing a single word of a prompt, I ask myself: "What exactly do I want to get from the AI?" For example, if it's about site planning, the goal might be: "Develop a landscape design for a 10-are plot, including a barbecue relaxation area, a small children's playground, a fruit orchard, and a decorative pond, taking into account an existing elevation difference of 1.5 meters." The more specific the goal, the easier it will be for me to evaluate the result.

2. Creating an Initial Prompt Hypothesis

Based on the goal, I draft the first version of the prompt. It includes key elements that the AI should consider. At this stage, I try to be as detailed as possible while avoiding redundancy. I imagine myself as a client articulating their wishes to a designer.

3. Iterative Testing and Refinement

This is the longest and most crucial stage. I run the prompt multiple times, making small changes each time and analyzing the results. For example, if I see that the AI consistently suggests paths that are too narrow, say 0.5 meters wide, even though a minimum of 1.2 meters is needed for a garden cart or stroller to pass comfortably, I adjust the prompt by adding size specifications. I check how the AI reacts to different phrasings, synonyms, and word order. It's like fine-tuning a precise instrument where every screw matters.

4. Variability and Edge Cases

Once the basic prompt starts yielding stable results, I begin testing it for variability and edge cases. What if the plot is not 10, but 6 ares, and also narrow and long? What if it's 20 ares and almost square? How will the AI react to a request for "minimalist design" compared to a "lush English-style garden"? What if the plot has a steep 10% slope or is located on the north side with little sun? My task is to ensure that the prompt works across the widest possible range of conditions. This allows me to guarantee that formspace.design will be useful for most plot owners, regardless of their individual characteristics.

Common Prompt Writing Mistakes and How I Avoid Them

Over time, I've identified several common mistakes that can "confuse" even the most advanced AI:

• Lack of specificity. Phrases like "make it beautiful" or "cozy" are uninformative for AI. I always try to translate such general wishes into specific parameters: "create a relaxation area with a pergola and fire pit," "use low-maintenance plants," "provide a drainage system with a slope of at least 2% for effective water runoff."
• Contradictory requirements. You cannot ask for "maximum open space" and simultaneously "dense forest on the plot." Such requests cause an "interest conflict" for the AI and lead to illogical results. I learn to formulate prompts so that they are internally consistent.
• Ignoring context. Forgetting to specify the climatic zone, soil type, or presence of utilities means depriving the AI of important information that influences the choice of plants, materials, and engineering solutions. I always include key contextual data in prompts.
• Too many details at once. Sometimes it's better to start with a general request, get a few options, and then gradually refine the details. This allows the AI to first grasp the overall concept and then elaborate on the nuances. As I wrote in the article AI tools I tested and rejected, choosing the right approach to formulation is half the battle.

Criteria for Evaluating AI Concept Quality

When the AI generates a concept, I evaluate it based on several key parameters, drawing on my experience:

• Functionality. Do the proposed zones match the stated requirements? Are they logically arranged? For example, the barbecue area shouldn't be too far from the kitchen, and the children's playground should be easily visible from the house. I check if paths, walkways (e.g., main path width 1.5 m, secondary 0.8 m), and areas are realistic.
• Aesthetics. Does the design match the chosen style? Is it pleasing to the eye? Here, my "artist's eye" and many years of experience with painting come into play.
• Feasibility. Is it possible to build this? Are elevation changes and existing utilities considered? It's important to me that the concept isn't a fantasy but represents a real project. For example, I always assess how difficult it would be to implement drainage or an automatic irrigation system, and whether these points are considered in the AI's proposal.
• Economic Viability. AI can suggest very expensive solutions. My task is to find a balance between the desired outcome and a reasonable budget, teaching the AI to propose optimal, not just the most impressive, options. This is especially important for plot owners who want quality design without excessive spending.
• Scalability. Does the prompt work equally well for plots of different sizes? From small 6-are plots to spacious 20+ ares, as well as for various shapes and configurations.

How to Plan Your Plot Using AI in formspace.design

When you open formspace.design, my goal is to give you a tool that has already undergone thousands of tests and is ready for use. Module 2.1 "AI Concept" is the heart of the idea. You upload your plot plan, specify its exact dimensions, orientation to cardinal directions, and existing objects – house, garage, large trees. Then you describe your wishes in detail: what functional zones you want to see (e.g., barbecue area, small vegetable garden, swimming pool), what style you prefer (modern minimalism, natural style, classical). The more precise and detailed your request, the more relevant and well-thought-out result you will get.

The system generates several concept options, considering the entered data: terrain, existing structures, and your preferences. You can choose the best option, refine it manually, or use it as a starting point for more detailed planning. For this, Modules 1.x "Plan/Zones/Objects" are perfectly suited, where you can precisely arrange elements, set the dimensions of paths and areas, and place plants and small architectural forms. For example, you can specify that the main path should be 1.5 m wide, and secondary paths 0.8 m, to ensure comfortable movement. Module 3.1 "Design by Photo" also helps quickly visualize ideas, and behind the scenes, this is the result of long iterations in selecting prompts to create photorealistic and stylistically consistent images. I discussed more about the AI's capabilities in our product in the article What AI in formspace.design can do now, and what's in development.

FAQ

Can one prompt be used for all plots?

No, this is highly inefficient. Each plot is unique in its characteristics — size, shape, terrain, insolation. Moreover, each owner has their individual wishes and style preferences. I strive to develop flexible prompts that can be adapted, but a general "ideal garden" formulation without specifics will not yield the desired result. It's better to use template prompts as a base and then refine them for specific conditions.

How quickly does the AI learn from my mistakes?

The AI itself does not "learn" from your individual mistakes in real-time like a human. As a developer, I collect and analyze user feedback, study which requests lead to unsatisfactory results, and based on this, I adjust the core prompts and request processing logic. This is a continuous, iterative process of system improvement that I control.

What if I don't like the AI's result at all?

First, try changing your request. Perhaps you used overly general formulations or missed important details. Add more specifics about the style, functional zones, or materials. If the problem persists, it's a valuable signal for me. You can always use the feedback button in formspace.design so I can analyze your request and understand what went wrong to improve the system.

How much does the AI consider the project budget?

The AI does not directly operate with specific monetary amounts, as the cost of materials and labor varies greatly. However, in the prompts, I incorporate the ability to specify "budget constraints" (e.g., "economy option," "mid-range budget," "premium class"). This allows the AI to propose solutions that, for example, minimize the use of expensive materials or complex structures. My task is to teach it to offer optimal solutions, not just the most impressive ones.

Can prompts replace a landscape designer?

Prompts and AI are powerful tools to assist and accelerate the work of a designer or plot owner. They can generate ideas, create concepts, and visualizations in minutes. However, the human factor — experience, creative thinking, deep understanding of context, ability to work with living plants, and personal communication with the client — remains indispensable. formspace.design is designed to augment human capabilities, not replace them.

How do you verify the realism of AI suggestions?

My many years of experience in landscape design, which began in 1997 and then with the founding of "Dom i Sad" (House and Garden) company in 2000, is key. I check each generated concept for feasibility: whether slopes for water drainage are accounted for, if the dimensions of zones and paths are realistic, how difficult it would be to implement drainage or an irrigation system. I often imagine myself on the construction site to assess practical aspects and ensure that the proposed solution is not only beautiful but also achievable.

My journey to creating formspace.design is a continuous process of learning and improvement. I believe that the future of landscape design lies in the synergy of human experience and artificial intelligence capabilities. Try formspace.design and see for yourself how carefully crafted prompts can transform your plot.

Related Materials

• Read more