River on AI-Assisted Programming

33.22 minutes · 452 sentences · River Cox + Adam Kane · Open in Fireflies

• Start with AI-built internal tools even when the end user does not use an LLM directly.
• Proposed practical targets: prefab and screenshot automation, Google Sheet ingestion, a Memento-aware debugging path, and safe isolated branches for experiments.
• Explicit governance: humans own the work; agents assist. Validation and review are often the expensive part.
• Adoption should spread through useful help and screen-sharing, not a mandate for “100% AI adoption.”

14:13–14:28 “I would like to see it be human done, AI assisted… Where I’m concerned is if it’s AI done, human assisted or just AI done period… with AI assisted programming, review takes more time…”

22:37 “I think… if we talk about 100% AI adoption, that might be a difficult thing to achieve.”

66.01 minutes · 1,111 sentences · 13 participants · Open in Fireflies

• River recommended Codex or Cursor to help William debug a current issue.
• He preferred Cursor’s debug mode because it forms hypotheses, asks for the exact logs needed, and confirms the issue before editing.
• His criticism of many coding agents: they guess from the description, then “change everything” without reproducing the problem.

43:50–45:01 “You can get Cursor and switch into debug mode… It documents and logs the issue and doesn’t fix anything until it confirms what the issue is… most coding agents… [guess]… and then it changes everything.”

Attendance note: Jonathan Cox and Keneth Vernon were both in the meeting. Jonathan briefly entered this AI-debugging exchange; Ken was an attendee, but the verified segment does not attribute an AI position to him.

64.28 minutes · 851 sentences · 24 participants · Open in Fireflies

• River demonstrated an internal coding assistant for ForgeFX standards and an AI-assisted Jira-to-code workflow.
• He used Cursor for repository-aware search, code changes, Unity YAML inspection, and iterative command-line-tool testing.
• He repeatedly kept a human checkpoint: review the approach, request changes, test the result, then accept or finish manually.

45:14–47:56 River contrasted plain ChatGPT with Cursor’s repository access and multi-step work, then showed Cursor searching, proposing a fix, editing code/YAML, and returning control for feedback and testing.

Attendance note: Jonathan Cox, Keneth Vernon, Miguel Whitney, Carl Lowther, Brandon Floyd, William Smith, and other ForgeFX production/development staff attended.

62.80 minutes · 872 sentences · 9 participants · Open in Fireflies

• River tested three approaches against the same current Halliburton bug: manual work, Cursor, and a more autonomous Dev Bot.
• Cursor found the correct class faster than River’s manual attempt; Dev Bot went further without detailed prompting.
• The rehearsal also exposed the reliability gap: Cursor did not finish and River reported a network problem. Adam recommended a prepared fallback for live demonstrations.

27:11–36:48 “I’m going to try to fix it three different ways… Cursor got me much further than I got on my own in the same amount of time… Cursor didn’t finish it… I’m having some kind of network issues with Cursor.”

Attendance note: Jonathan Cox attended, but the verified AI segment contains no substantive statement from him.

147.35 minutes · 825 sentences · 15 participants · Open in Fireflies

• River openly described the in-progress system as “50% AI generated and 50% vibe coded,” welcomed concerns, and offered to help Ken evaluate the work.
• He identified performance and complexity debt in interacting terrain systems that required simplification and refactoring.
• His model of agent productivity: spectacular during a small project’s honeymoon phase, then sharply weaker as requirements and context accumulate.

26:22–26:48 “There’s a honeymoon phase where it is just kicking ass… doing like a hundred times more than I could… then as the project gets bigger… it starts to choke… most of our AI generation stuff, they don’t have long term memory.”

Attendance note: River’s remarks were directed to Ken and acknowledged Ken’s concerns, but the retained evidence does not contain Ken’s underlying comments. Jonathan also attended without a substantive retained AI statement.

82.66 minutes · 1,209 sentences · River Cox + Adam Kane · Open in Fireflies

• River described frustration with AI tooling that looks 90% complete but whose last 10% takes months, years, or never finishes.
• He was personally using AI heavily, including Cursor, and said TypeScript’s fast compilation helps expose hallucinations quickly.
• He saw web/TypeScript/Node workflows as substantially easier for AI agents to test than Unity’s slower, manual build-and-runtime path.
• He preferred measurable outcomes over flashy demonstrations and found a specialized Unity tool more useful than the Cursor Unity MCP at that time.

48:38–48:51 “TypeScript from AI is amazing because the compilation is fast and the compilation fixes hallucinations… Everything I’m doing for the license generator in Unity and outside of Unity is AI…”

51:43 “I don’t think it has to be so flashy. I think just a clear and consistent and measurable outcome from AI.”