The framing that actually matters here isn't "is AI good or bad for coding" — it's what kind of collaborator it is. Treating any current LLM as "always right" or "not worth using" are both wrong in the same way: they skip evaluating what it's actually consistently good at versus where it needs a second look, task by task.
What holds up in practice for Unity work:
Debugging with full context. Pasting the exact error plus the relevant script gets a specific diagnosis (e.g., a missing GetComponent() call or an unassigned Inspector field) instead of generic troubleshooting. Bare error text alone gets you a much weaker answer.
Open-ended review over yes/no questions. "Review this NavMeshAgent patrol script for bugs and mobile perf issues" surfaces things you didn't think to ask about — a GetComponent() sitting in Update() instead of cached, a missing null check. A yes/no prompt just answers the narrow question you asked.
First-draft content, not final content. Design docs, item flavor text, NPC dialogue — genuinely fast for a first pass, needs a real editing pass for voice and lore consistency before it ships.
What doesn't hold up: anything version-sensitive on very recent API changes without explicitly having it search first, and using it as your only reviewer on a team project rather than a first pass before a human looks at architecture and conventions.
Full prompt examples and a fair comparison against Copilot/ChatGPT for different parts of a dev workflow: https://digitaltoolify.blogspot.com/2026/06/claude-ai-for-game-developers-practical.html
Curious what other devs are actually using this for day to day beyond debugging — anyone building it into an actual CI step or internal tool via the API?
Top comments (0)