You know what happens when you let AI design a game’s entire character progression system?
Well predictably, it tries to fabricate brand-new mechanics and explain how each of them works to you. In this experiment, we decided on the concept of a rogue-class character and proceeded to feed it with prompts.
The results revealed something crucial about AI’s current capabilities. The system itself, on paper, and at a glance, looks remarkably comprehensive and professional looking. However, the gap between “looks good on paper” and “actually works in practice” keeps itself at a good distance that only humans can bridge for now.
Why Skill Trees Are So Hard to Design (And the AI Temptation)
Skill trees are notoriously difficult to balance because they are not just lists of playstyle gimmicks. Outside of a gaming perspective, you can view these things as motivation nodes, that require in-lore logic, stat balance, diversity (among other skill trees), and preferred playstyles. As one designer at GDKeys put it, “A good skill tree forces the player to make impacting, committing, long-lasting choices for the development of his playstyle.”
Needless to say, traditional skill tree design is worked from the ground up (pun not intended). Developers spend months hand-tuning progression curves, ensuring no single path becomes overpowered while avoiding “dead branches” that feel useless. Each ability needs to feel meaningful, prerequisites must make thematic sense, and the whole system has to scale properly from the beginning up to endgame builds.
With generative AI having entered the automated data production scene over the last few years, many people have dabbled with the idea of relegating the delicate responsibility of building skill trees to such “artificial developers.” Think about it. AI is perfectly suited for the systematic thinking needed for writing skill trees. They even excel at pattern recognition, and can follow rules properly so long as they have a solid foundation.
In addition, once it has the initial variables, it can iterate almost instantly. You can even suggest thematic connections between abilities that you fancy, with the constant variable reminding you of mathematical balance. It is a potential boon for time-pressed developers and teams of very small sizes! Like, why spend weeks crafting skill progressions when AI might generate dozens of options in minutes?
But as we already concluded in the intro, all of these theoretical efficiency gains only matter if the output is really usable. And this is where the real experiment begins.
Prompting Our Way to a Skill Tree
This experiment is specifically done using Claude 4.0 Sonnet, the latest balanced LLM of Anthropic. This particular model is free to access even without a subscription, albeit with a limited number of prompts per specified period of time. For our “test subject,” we went for a classic fantasy rogue character, and decided on three progression branches: stealth, poison, and traps.
Then, we structured our experiment around five increasingly complex and progressive prompts (each building on top of one another). We believe that this is somewhat similar to how developers might use dedicated AI skill tree generators for rapid prototyping in an official capacity.
By the way, you can try doing the same prompts to the exact same LLM (and the same version) as well. Check out what kinds of different results you can get!
Prompt 1: “Design a complete skill tree for a rogue character in a fantasy RPG. Include three main branches: stealth, poison, and traps. Each branch should have 4-5 tiers with 2-3 skills per tier.”
Result of Prompt 1: Claude delivered far more than we expected. What we expected to be a simple progression outline, turned out to be a comprehensive system with detailed skill descriptions, prerequisites, and even usage limitations. The stealth branch alone included abilities like “Shadowstep” (instant 15-foot teleportation) and “Phase Step” (40-foot teleport with bonus damage), each with specific cooldowns and mechanical interactions.
What impressed us was the thematic coherence. The AI understood that stealth abilities should build upon each other. For example, basic movement skills unlock advanced positioning techniques. Learning those skills then enabled master-tier abilities like “Umbral Form” (temporary incorporeality). On paper, the progression felt logical, with each tier introducing meaningfully more powerful capabilities.
Prompt 2: “Now balance these abilities across the three tiers, ensuring no single skill is overpowered. Include synergy bonuses between different branches.”
Result of Prompt 2: Here, Claude showed unexpected sophistication, implementing a skill point economy (5/10/15/20/25 points per tier) and creating cross-branch synergies like “Shadowtoxin,” which are explained as poisoned weapons that don’t break stealth and gain unresistable damage. Seems that the AI was really thinking systematically about player builds and mechanical interactions.
Prompt 3: “Identify potential balance issues in this skill tree and suggest fixes. Which skills might be too weak or too strong?”
Result of Prompt 3: Our third prompt tested the AI’s critical thinking by asking it to identify balance issues in its own creation. Claude caught several obvious problems. The skill “Death’s Embrace,” in particular, deals 25% max HP per round would indeed trivialize most encounters. Unfortunately, looking at the easiest-to-spot ones was the limit of its curating capabilities. The AI could spot mathematical outliers, but couldn’t anticipate how players might exploit subtle interactions between systems.
Prompt 4: “Generate 3 alternative level 20-character builds using this skill tree, focusing on different playstyles.”
Result of Prompt 4: The response to this prompt generated three complete character builds with exact point distributions and damage calculations. The “Phantom Assassin” build allocated 70 points to stealth and 30 to poison for maximum burst damage (averaging 14d6 per attack). These builds were mathematically sound and thematically distinct, demonstrating that AI could understand not just individual abilities but how they combined into coherent playstyles.
Prompt 5: “Create a completely new, unique mechanic for this skill tree that hasn’t been seen in traditional RPGs.”
Result of Prompt 5: Needless to say, the real potential surprise should be with this last prompt. And its answer? The “Shadow Memory System.” It explained it as a mechanic where rogues collect “Shadow Fragments” by “observing” enemy abilities, then spend them to counter or replicate those same abilities. Creative points where it is due, of course, plus the application of the “knowledge-as-currency” concept for its proposal. However, since it looks like the LLM only answered this question partially in a vacuum, it seemed to never consider the implementation and balancing aspect of the skill.
Verdict: Impressive, but Might be Hollow
The more we examined these outputs, the more questions emerged. The poison stacking system created elegant mathematics but might reduce complex boss encounters to waiting games. The trap networks sounded tactically sophisticated but could break encounter pacing if players could prepare extensively. Ironically, the biggest probably was probably the sheer comprehensiveness, as Claude generated enough content to fill the entire space of this article several times over.
AI in the Real Game Development World
This little skill tree experiment reveals a broader truth about AI’s role in game development: “apparent” complexity. We have proven beyond a doubt that it is very good at generating sophisticated, comprehensive content that follows established patterns. However, as clearly shown by its flaws, distinguishing genuinely good design from convincingly presented ideas still requires human expertise.
Despite this, AI tools are still being adopted for specific tasks. Even if creative decision-making remains firmly in human hands, these tasks still benefit from better spontaneity while being automated, thanks to generative AI’s specific quirks. AI character creation companies like CandyAI could be the next steps into using new generative models to materialize an in-game character based on the newly created skill-trees, at least for that initial game-setup wizard. You could just describe in text what you’re looking for, and how would you balance the skills, and that should be it, you hit “Create” and – poof, here it is.
Industry Examples
Ludo.ai’s Game Concept tool serves as an AI-powered co-writer, generating text and images for mechanics, story, and characters, while Magic Media studio reports using AI tools for procedural content, game art, and quality assurance workflows. The A16Z Game Developer Survey found that 70% of respondents are either using or plan to use 3D AI tools, with many exploring AI for runtime content generation. Proliferation of the technology, at least economically, is already beyond its first steps.
The most successful implementations focus on specific, well-defined tasks. Middle-Earth: Shadow of Mordor’s Nemesis System allows NPC enemies to remember their interactions with the player, creating dynamic rivalries that feel personal. FIFA’s “Player Personality System” uses AI to give each virtual player a distinct identity, with unique behaviors and tactical decision-making.
Human Curation for the Foreseeable Near Future
“We view AI in game development as another tool in our arsenal, one to be used judiciously, explored for its capabilities, and applied effectively, rather than as a one-size-fits-all solution,” explains Magic Media’s team. This perspective acknowledges that AI is indeed a powerhouse in its expertise, while still recognizing where it falls horribly flat.
You see, even with just five mere prompts, our skill tree experiment already shows exactly why human oversight is still necessary. AI excels at pattern recognition and can generate content that follows established rules. But, it struggles with “weaving the entire narrative,” something that human developers naturally do without even actively thinking.
While they created abilities that sounded impressive and followed logical progressions, a lot of the more practical aspects remain largely untested. How would it interact with the larger game system? Is it fun to play? Is it fun to play against? These things still require an intuitive understanding of player psychology, a critical context that generative AI currently lacks.
AI Builds the Scaffolding, You Build the Tower
This points to AI’s emerging role as a sophisticated brainstorming partner rather than a replacement designer. AI-powered development tools, for instance, can cut production time at the conceptual stage, allowing teams to iterate faster with an already workable framework rather than being stuck for a while from scratch. AI and human creativity combined as a collaborative effort rather than a competitive element.
Turning any of AI’s concepts into actual gameplay might require extensive human curation, testing, and refinement. Still, that sweet spot where artificial intelligence amplifies human creativity rather than replacing it makes all those production nudges more detailed, inviting better insight, and perhaps even newer branching ideas.
Indeed, we’ve witnessed Claude create multiple valid interpretations of what a rogue progression system could be. The Shadow Memory mechanic, albeit wonky in practice, shows that “creativity” via refreshed concepts is achievable even with current generative AI. But shaping that material into something that feels meaningful, balanced, and genuinely fun? That’s still distinctly human work. At least for the next few years or so.
