Kyle Kukshtel

Kyle Kukshtel is a game developer and toolmaker who created Dreamwalker, an interface for exploring image-generation models through branching visual experiments. His subsequent AI work includes using coding agents to update production software and testing where model-generated game design breaks down. He uses the handle kkukshtel on Twitter and GitHub.[3][4][5][10][9]

Kukshtel's personal website, as of September 2026, lists him as a senior developer on Unity's runtime team working on IL2CPP and the Mono-to-CoreCLR transition. His About page also identifies him as Afterschool Studio's lead developer, creative director and co-founder.[2][3]

Kukshtel developed Dreamwalker for the first NYC AI×UX meetup after finding image-generation tools too focused on constructing “the perfect prompt.” He wanted a more playful process resembling a moodboard: collecting possibilities, following promising directions and pruning branches. In his October 9, 2023 announcement, he described generated images as starting points that could help users overcome the blank page and find their next step.[5]

The earlier meetup demonstration used a DALL·E interface built in Unity. It represented the path through prompt space as a graph, with the application generating prompts while the user guided the direction of exploration. Kukshtel demonstrated it at NYC AI×UX; Paul Butler described the demo in a recap published by Latent Space on June 1, 2023.[1]

After the presentation, Kukshtel set the project aside while considering a larger feature vision. He ultimately chose to release its core features on October 9, 2023. His stated aim was to encourage people to “think outside the chat prompt for image model interaction” and inspire other unconventional AI tools.[5]

The public release documentation describes a later version offering SDXL via Replicate or a user-selected Glif workflow. Images appear as nodes on a canvas that users can zoom and pan. A plus button spawns another image using a selected transform: Vary resubmits the same prompt, while Sublime and Weird change its semantic direction. Users can add named transforms or enter a one-off instruction that modifies the current prompt. These controls let users branch from an existing result without composing each new prompt from scratch.[6]

Kukshtel used his moodboarding website as a production testbed for AI coding. In his April 18, 2025 account of its Premium Update, he reported releasing a Svelte rearchitecture, customization options and payment support. He chose Svelte to replace several large JavaScript files with smaller, composable units, making the code easier for him to understand and reducing unrelated code entering model context. Sonnet 3.5 performed the migration through Cline; he later switched to Claude Code, preferring its terminal workflow and context handling. He qualified his impression of better coding quality because the switch coincided with an upgrade to Sonnet 3.7.[10]

A gallery bug illustrated why he still needed to understand the generated code. Both Cline and Claude Code misdiagnosed images loading without a properly initialized gallery. Kukshtel traced the problem to loading-state changes causing Svelte to rerender a conditional template and remove the instantiated gallery. He also argued that choosing features remained a product-design responsibility: an agent's suggestion to add image editing, for example, missed the purpose he had set for the moodboarding app.[10]

In an April 30, 2025 experiment, Kukshtel tried turning the AI 2027 report into a browser strategy game inspired by Twilight Struggle, Imperial Struggle and Daybreak. He used Gemini 2.5 for rules, sprint plans and unit tests, then passed the plans to Claude Code for implementation. He stored project knowledge in CLAUDE.md and used Supabase and Puppeteer MCP integrations. He found the planning handoff useful, but remained uncertain about the effectiveness of browser tests and encountered difficulties testing against a local Supabase database.[9]

The deeper failure was in the rules. Kukshtel found that players could effectively win with their first hand by converting all their cards into operations points and advancing the AGI track to its maximum. He judged Gemini's design an unbalanced average of its references, although he liked its thematic event-card ideas. He concluded that he should design the game himself before generating an implementation plan, while retaining models for content ideation. He ended the account by saying he was stopping work on the unfinished prototype, warning that the apparent ease of prompting could encourage spending longer on a failing approach than direct work would require.[9]

In “UNSTABLE SCULPTURES,” originally published June 12, 2025, Kukshtel predicted that stabilizing diffusion-model output could become a major frontier for AAA game development. He proposed combining generated visuals with an underlying simulation to preserve consistent game state. Possible supporting tools included scene annotations, feedback from generated images, reuse of generated audio and buffering output to correct errors before displaying it. He distinguished these audiovisual possibilities from games where exactness matters, such as real-time strategy games and MOBAs. His forecast of console hardware for local inference was explicitly offered without insider information.[8]

His February 21, 2026 essay “On-Demand Software” explored what improving first-pass generation might mean for programming. “Programming languages are for humans,” he argued, connecting the idea to his Unity IL2CPP work: converting C# intermediate language into C++ intended for compilation rather than human editing. He speculated about interfaces for exploring the many programs implied by an initial request, including branching graphs of possible designs, and about custom software generated only for the duration of its usefulness. Direct binary generation and alternative interfaces to inference were proposals, rather than systems he claimed to have built.[7]

Before these 2025–2026 coding experiments and forecasts, Kukshtel had worked on game-authoring tools and volumetric media. He co-founded Scatter and worked on Depthkit, including its Unity import pipeline around a Kinect-based volumetric capture toolkit. This earlier work supplies practical context in building tools that connect creative material to game-engine workflows.[3][4]

On August 15, 2023, Kukshtel announced that his team's tactical-strategy game Cantata had reached version 1.0 from Early Access, describing his own involvement as nearly a decade. The game combined 4X, grand strategy, tactics, automation and RTS elements without separating tactical and strategic views. He built its runtime level, trigger and logic editor; the team used the same editor to create campaign maps.[11][4]