Shayne WinterAI Systems Lab

HomeAI systemsNeural Commander

AI-directed product development

Working playable prototype

Turn hands-on playtesting into an explicit, testable product system.

Neural Commander is a deterministic isometric strategy-game prototype. More importantly, it is a working example of how I direct AI-assisted development: observe the product in use, identify the real failure, define the system contract, coordinate bounded implementation, and require native runtime evidence before accepting the result.

Stack: C# · .NET 8 · GodotRole: Product direction, playtesting, system contracts, acceptanceMaturity: Playable prototype under active development
For a hiring managerThis project demonstrates product ownership across a complex stateful system: translating subjective play experience into precise behavior, keeping many subsystems consistent, directing AI-assisted implementation, and carrying active validation targets forward as explicit engineering work.
Neural Commander prototype with a deployed command structure, factory, units, resources, objectives, minimap, and command interface
Accepted native gameplay evidence from a working prototype checkpoint, showing the integrated interface and core gameplay loop.

The product challenge

A strategy game is a network of promises, not a collection of features.

Movement affects combat. Economy affects production. Deployment affects construction. Terrain affects navigation and visibility. The interface has to explain state without becoming the authority for it. When one subsystem changes, the rest of the product must remain deterministic, understandable, and testable.

That made the work a useful proving ground for the same systems discipline required in operational software: explicit contracts, stable authority, versioned data, failure reproduction, and acceptance based on behavior rather than intent.

What the prototype includes

A connected game loop rather than a scripted mockup.

Build and operate

Deployment, construction, resource extraction, production queues, unit control, docking, and recovery.

Compete and survive

Movement, combat, hostile behavior, waves, defensive decisions, visibility, minimap state, and host-authoritative outcomes.

Admit content safely

Versioned data packages, strict validation, deterministic loading, immutable state views, and controlled presentation assets.

How I directed the work

Convert “this feels wrong” into a contract an engineer can build and a test can reject.

Hands-on product discoveryPlayed the real build, compared intended and observed behavior, and identified the point where the system—not merely the screen—was wrong.
Requirements translationTurned observations into explicit simulation rules, data semantics, UI expectations, failure behavior, and acceptance criteria.
Architecture decisionsKept the simulation host authoritative, separated state from presentation, versioned content admission, and preserved deterministic replayable behavior.
AI workstream directionUsed specialist agents for bounded implementation and analysis while retaining tradeoff, correction, integration, and final acceptance authority.

A representative correction

One word can reveal the wrong system model.

During economy work, the implementation treated Alloy as physical cargo when the intended design treated Alloy as an economic resource and Mass as the physical load. That was not a label defect. It affected gathering, storage, transport, production, and user expectations. I corrected the domain model first, then required the implementation and tests to follow the corrected meaning.

That pattern—finding the conceptual error behind a visible symptom—is central to my work across games, business systems, embedded devices, infrastructure, and AI platforms.

Verification

Use the evidence layer that can actually answer the question.

  • Deterministic simulation and application tests protect rules and regression-sensitive state transitions.
  • Strict package and content validation rejects incompatible or malformed data before it enters the game.
  • Native Windows captures show actual gameplay, layout, scaling, selection, production, gathering, docking, and combat behavior.
  • Audio analysis and runtime diagnostics separate generated assets from what the executable actually renders or plays.
  • Hands-on playtesting remains the acceptance layer for usability, visual hierarchy, pacing, and whether the product feels coherent.

Current state

A substantial working prototype in active development.

Earlier checkpoints established broad deterministic coverage and native gameplay evidence across the core loop. The current phase is focused on later AI and defense behavior. Focused checks pass for the corrected work; three longer scenarios are the active validation targets before the next full current-source regression.

The present maturity is a working, heavily tested prototype. Release candidacy follows completion of those scenarios and a fresh full regression.

Transferable capabilities

Why the work matters outside game development.

  • Technical product direction
  • User acceptance testing
  • Complex-state modeling
  • Host authority
  • Deterministic simulation
  • Data contracts
  • Content validation
  • Regression design
  • Visual evidence
  • Failure reproduction
  • AI-assisted delivery
  • C# / .NET / Godot
Next AI case studyPegasus Cortex