3 Prompting Strategies to Guide AI Toward Original Solutions | With Rocket Thruster Case Study

A Bold Experiment That Changes How We Use AI

When using AI for development assistance, most people have settled into a fixed workflow—state your requirements, let AI assess feasibility, then AI recommends solutions based on existing industry practices. But one indie developer posed a bold question: Since you're an AI, why can't you break free from conventional human approaches and solve problems your own way?

This seemingly simple question actually touches on a deep issue in current AI applications—is AI merely a porter of human knowledge, or a tool with genuine innovative capabilities?

Whether AI possesses true creative ability remains debated in academia. Supporters cite "combinational creativity" theory, arguing that most human innovation is essentially recombination of existing knowledge, and AI has a natural advantage here—it can perform high-speed combinatorial searches across knowledge spaces far exceeding human cognitive bandwidth. Critics point out that AI lacks "transformational creativity"—the ability to fundamentally change conceptual frameworks.

Cognitive scientist Margaret Boden's classification framework from her classic work The Creative Mind provides valuable reference for understanding AI innovation boundaries—current large language models excel at combinatorial innovation rather than paradigm-level breakthroughs.

The Standard AI-Assisted Development Pattern: Why Mature Solutions Aren't Always Enough

This developer shared his workflow when using AI for development: before implementing a feature, he first has AI evaluate it on two dimensions—whether the feature is necessary and whether it can be simply implemented from a technical perspective.

AI typically provides quite professional assessments, but with one obvious characteristic: its suggestions almost always revolve around existing industry solutions. Whether it's technology selection, architecture design, or specific implementation, AI tends to recommend widely validated mature solutions.

This tendency is closely related to the training mechanism of large language models. LLMs learn the statistical distribution of human knowledge through massive text data—solutions that appear most frequently receive the highest generation probability. This phenomenon can be understood as a mild form of "mode collapse": the model tends to converge on the most common patterns in training data.

For example, when you ask AI how to implement frontend state management, it will most likely recommend Redux or Zustand, because these solutions appear most frequently in training corpora. This isn't AI lacking "creativity"—its default generation strategy is to maximize output likelihood probability, giving the "safest" and most "mainstream" answers.

Mature solutions mean stability, reliability, and community support, which is perfectly fine in itself. But for indie developers, sometimes what's needed isn't "the safest choice" but rather "the most innovative solution for the current scenario." This is precisely where prompt engineering needs to level up.

Breaking Convention: Using Prompts to Guide AI Toward Original Solutions

So this developer tried something interesting: he explicitly asked AI not to follow conventional approaches, but to solve problems using self-created methods.

His prompting strategy was straightforward—"Use your own self-created method, don't follow conventional approaches." The essence of this prompting approach is attempting to remove AI's "comfort zone" limitations, forcing it to escape the highest-frequency solution patterns in training data.

From a technical perspective, Prompt Engineering has evolved from simple instruction writing into a systematic technical discipline. Its core principle is guiding models to explore different regions of their latent knowledge space through carefully designed input text. Conventional prompts activate high-probability output paths in the model, while constraint-based prompts like "don't follow conventional approaches" essentially adjust the model's sampling distribution—reducing the output weight of high-frequency solutions and forcing the model to explore low-probability but potentially more innovative output regions.

This works similarly to adjusting the model's temperature parameter, except it's achieved through natural language rather than parameter tuning.

The developer also acknowledged that this is still in an experimental stage—he simply wanted to test the boundaries of AI's potential. This spirit of exploration is itself commendable—only by continuously probing AI's capability boundaries can we discover new application possibilities.

A Real AI Innovation Case: Generative Design of Rocket Thrusters

Having AI create its own solutions isn't pure fantasy. The developer mentioned a real case: someone used AI to design a rocket thruster, and the structure AI designed was in a form incomprehensible to humans.

This case comes from AI applications in engineering design, commonly known as Generative Design or topology optimization. Given constraint conditions (weight, strength, materials, etc.), AI generated a structure through extensive iterative calculations that completely defies human intuition but is entirely physically viable. These structures often look like the skeletons of organic organisms, filled with irregular holes and curved surfaces, yet their mechanical performance surpasses traditional human engineering designs.

Representative generative design tools include Autodesk's Fusion 360 and nTopology. The technical principles are typically based on topology optimization algorithms (such as the SIMP method—Solid Isotropic Material with Penalization) or evolutionary algorithms, searching for optimal material distribution within a given design space through thousands or even millions of iterations.

Unlike traditional human design workflows, generative design doesn't start from preset geometric shapes but from mathematical descriptions of physical fields (stress fields, thermal fields, etc.), letting algorithms "grow" the optimal structure on their own. NASA's Jet Propulsion Laboratory (JPL) used this technology to design support structures for interplanetary probes, with the final solution being 35% lighter than human engineers' designs while maintaining higher strength.

The reason these AI-generated structures look like biological bones is that natural bones themselves are the result of millions of years of evolutionary optimization—AI completes in hours what evolution took millions of years to achieve.

This demonstrates a key fact: AI's "way of thinking" differs from humans, and when we allow it to break free from human design paradigms, it may find solution paths humans never imagined.

Three Prompting Strategies: Guiding AI Beyond Conventional Thinking

Based on the experiments and cases above, we've summarized three practical prompting strategies to help you unlock more innovative possibilities in AI-assisted development.

Strategy One: Exclusion Method—Explicitly Tell AI What NOT to Do

Most people using AI coding assistants habitually ask AI to follow "best practices." But this approach may limit AI's output space.

The exclusion strategy has a corresponding theoretical basis in cognitive science. Psychological research shows that "functional fixedness" is one of the main barriers to human innovation—people tend to see only the conventional uses of objects while ignoring other possibilities. While AI doesn't have human cognitive biases, the high-frequency patterns in its training data produce a similar "digital functional fixedness." By explicitly excluding specific solutions, we're effectively helping AI overcome this statistical-level fixedness, forcing it to explore more distant node connections in its knowledge graph.

Try telling AI not to use specific frameworks or patterns and see what alternatives it proposes. For example:

"Don't use React's conventional state management solutions. Give me a lightweight alternative approach you designed yourself."

The core of this type of prompt is using "exclusion" to open AI's creative space—first tell it what not to do, then see what new ideas it comes up with.

Strategy Two: Constraint-Driven Method—Set Only Boundary Conditions, Don't Preset Implementation Paths

Only describe the goal you want to achieve and the constraints, leaving "how to do it" entirely to AI. For example:

"I need to complete this calculation within 50ms with memory usage under 10MB. You decide what method to use."

The constraint-driven method is rooted in mathematical optimization theory. In operations research and engineering optimization, a classic insight is: good constraint definitions are more important than good initial solutions. When we only set boundary conditions for AI (such as performance metrics, resource limits) without presetting implementation paths, we're essentially defining a high-dimensional feasible region, letting AI freely search within this space.

This aligns with the concept of "reward shaping" in reinforcement learning—define what constitutes a "good result" and let the agent discover strategies to achieve the goal on its own.

This approach gives AI maximum freedom and often produces unexpected technical solutions. It follows the same logic as generative design of rocket thrusters—define boundaries, let AI find the optimal path itself.

Strategy Three: Multi-Round Iteration Method—Progressively Refine Through Follow-up Questions

First have AI propose unconventional solutions, then gradually assess their feasibility, continuously refining solution details through follow-up questions. This progressive prompt engineering approach preserves creative space while maintaining quality control over the final solution.

In practice, this can be broken into three steps:

1. Round One: Free Divergence—Have AI propose multiple unconventional ideas
2. Round Two: Deep Evaluation—Follow up on the most promising solutions with technical details and risk points
3. Round Three: Implementation—Request complete implementation plans and considerations

This multi-round iterative methodology actually simulates the classic "Double Diamond" model of divergence-convergence in Design Thinking. The first round corresponds to the divergence phase, encouraging as many creative ideas as possible; subsequent rounds gradually converge, filtering out truly viable solutions through technical validation and risk assessment. This structured exploration approach is more likely to produce high-quality innovative solutions than single-shot prompting.

Which Scenarios Are Suitable for Letting AI Freely Explore

It's important to note that AI's "self-created methods" aren't always reliable. In production environments, the stability and maintainability of mature solutions remain crucial. The following three types of scenarios are more suitable for this exploratory approach:

• Personal projects or prototype validation phases: Low cost of failure, allowing bold experimentation
• When existing solutions genuinely can't meet special requirements: When conventional paths don't work, let AI find alternative routes
• Pure technical exploration and learning processes: The focus isn't on whether results are usable, but on broadening perspectives

Redefining AI's Role: From Knowledge Retrieval to Innovation Engine

Although this indie developer's experiment is still in early stages, it points to a direction worth deep consideration: Are we over-relying on AI as a "knowledge retrieval engine" while overlooking its potential as a "creative thinking tool"?

When AI has been trained on massive amounts of human knowledge, it can not only recall this knowledge but should theoretically be able to combine and break through on the basis of it. The key is whether we give it this opportunity through appropriate prompts. As the generative design case of rocket thrusters shows, sometimes the optimal solution lies precisely in the blind spots of human cognition.

From a broader perspective, this transformation reflects the evolution of human-AI collaboration models. Early AI applications were primarily a one-way retrieval model of "humans ask, AI answers"; but when we begin using prompts to guide AI in creative exploration, the human-AI relationship is shifting toward a collaborative model of "humans define the problem space, AI explores the solution space."

In this model, humans' core value is no longer mastering specific technical solutions, but rather asking the right questions, defining appropriate constraints, and judging the feasibility of AI's output—which is precisely where human intuition and experience remain most irreplaceable.

Next time you use AI for development assistance, try saying: "Don't follow conventional approaches—use your own method." You might harvest some unexpected inspiration.

Key Takeaways

• In AI-assisted development, AI typically tends to recommend existing mature industry solutions, directly related to how LLMs maximize output likelihood probability, which may limit creative space
• Explicitly requiring AI to break from convention and use self-created methods essentially adjusts the model's sampling distribution through natural language, exploring low-probability but more innovative output regions
• The real case of AI-designed rocket thrusters proves that generative design based on topology optimization can find superior solutions when breaking free from human design paradigms
• Three prompting strategies—exclusion method, constraint-driven method, and multi-round iteration method—open AI's creative space from different dimensions
• This exploration is currently more suitable for prototype stages and personal projects; production environments still require weighing stability
• The core shift in prompting strategy: from presetting implementation paths to only setting goals and constraints, driving human-AI collaboration from one-way retrieval toward collaborative exploration