# AI, engine and asset choices Checked against the linked official sources on 2026-10-25. This is a practical starting guide, not a head-to-head benchmark. Availability, names, plans and limits can change. The creator workflow should survive those changes. ## Pick the application before arguing about model names Confirm whether your app can access a project folder, edit files, run the game and inspect evidence. Do not assume a model's advertised capability is enabled in every app or subscription. Ask for missing capabilities to be named. Start with one main assistant and a small project. Add a second assistant when it has an independent job or a review question. ## Dated model examples - OpenAI's current guide positions GPT-6.1 Sol as a balance for complex coding work, GPT-6 Astra for the most demanding work, and GPT-6 Luna for focused faster/lower-cost tasks. A reasonable KCS trial is the balanced option for a small build, then compare a stronger option on one unresolved problem. This is our workflow suggestion, not proof of measured savings or a universal ranking. Source: https://developers.openai.com/api/docs/guides/latest-model - Claude Code's model aliases position Sonnet for daily coding, Opus for complex reasoning and Haiku for simple efficient tasks; its best/Fable options are account-dependent. Check what the alias resolves to in your own account. These are vendor descriptions, not KCS comparative test results. Source: https://code.claude.com/docs/en/model-config - Google's catalog lists Gemini 4 Argon as its flagship reasoning model and Gemini 3.8 Flash for software engineering and agentic work. If those options are in your tool, compare on a bounded planning, review or implementation task rather than switching a whole project on a promise. Source: https://ai.google.dev/gemini-api/docs/models A stronger model may help when dependencies are confusing; a faster model may be adequate for a clear bounded job. If the same failure repeats, improve the evidence and scope before paying for more attempts. Browser agents and model APIs may bill differently. Check your actual plan; no paid AI subscription is included in this course. ## Free development routes - Browser route: have the AI build a small HTML/CSS/JavaScript game that runs in your browser. No separate game-engine purchase is needed; editor/AI/hosting services may have their own costs. The AI should handle files and give exact run steps. - Godot: a free/open-source engine under MIT. Engine redistribution carries notice/license requirements. It is a separate deeper track, the dedicated Godot track uses separately recorded example-project evidence. Source: https://godotengine.org/license/ - GDevelop: its official Free offering provides an open-source engine and limited included publishing/services. Check current plan limits; cloud/AI/convenience services can be paid. It is an alternative to investigate, not a claim this course was tested in GDevelop. Source: https://gdevelop.io/pricing Choose the engine for the project and available workflow. Do not change engines every time an ordinary bug appears. ## Free assets that save art-production time Start with a coherent set rather than twenty unrelated downloads. Ask the AI for a small asset shopping list: player, environment, collectible, hazard, interface and a few sound cues. Prototype with shapes when art would slow the first loop. Kenney's support page says assets on its asset pages are CC0 and usable commercially; read the specific downloaded pack's included license and do not use Kenney's reserved logo. Source: https://kenney.nl/support OpenGameArt entries have different licenses. Read the individual entry and its requirements; “free download” does not settle attribution or redistribution permissions. Source: https://opengameart.org/content/faq Keep an asset register: title, creator, original URL, license/version, download date, attribution text, modifications and whether redistribution with source is permitted. Do not package someone else's asset collection as your own KCS store product. Ask: “Find three coherent options from official creator pages, show licenses and credit requirements, and wait before downloading or integrating.” Let the AI handle resizing/importing only where your actual tools permit it. Third-party assets are suggested, not bundled into this course. No engine or model vendor endorses this KCS course. ## Avoid prompt and workflow bloat Teach the AI a small working agreement, then keep the current ask clear. Do not paste every workflow rule into every message. OpenAI's guidance warns against bloated context and advises calibrating task-specific delegation and tests. Sources: https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra https://developers.openai.com/api/docs/guides/latest-model ## Team work Parallelism helps when jobs can be done independently: one AI audits readability, another explores licensed art options, and the builder implements the agreed mechanic. Extra agents can increase cost and coordination. Use one integration owner; assign bounded files or separate copies; require a short handoff and actual evidence. Never ask two agents to edit the same live files blindly. If your app has no subagents, use separate chats with copied briefs and manual handoffs. That is a valid small-team workflow, not automatic coordination.