I wrapped up packaging for Nucleate, focusing on smoothing out FFmpeg and Ollama installations. The Windows setup went pretty smoothly, but the Mac install failed on non-admin accounts. Instead of trying to bypass those security restrictions, I decided to improve the error messages to make the problem clearer. I also spent time testing different app versions, backend setups, and hardware to stabilize things, fixing bugs and improving the first-time install flow—especially around handling installs for home root versus non-home root users.

After several iterations, I released Nucleate version one on Gumroad. The traction has been minimal so far, with just a few clicks coming from Reddit and Discord. I’ve been trying to promote it on Discord and Notion, but it’s mostly been silent. I also got some harsh feedback about the typing speed claims I made—advertised as 40–50 WPM but apparently more like 4–8 WPM in reality. I adjusted the messaging to clarify that those numbers reflect drafting and editing speeds, not raw typing speed, but I’m not sure if that helped perception much. I’m thinking about trying the personal knowledge management subreddit’s monthly self-promo space next to see if that gets more attention. I’m also experimenting with targeting students by offering a free lecture transcription tool to get some organic engagement.

On the AI side, I started working on Godex, an agentic AI assistant for open-source game development. I built a basic Qt app using Ollama’s API that can answer questions about projects and scenes. The pipeline crawls Godot’s documentation, chunks it, and embeds it for local RAG. The embedding process took hours but worked, and Ollama’s Qwen LLM uses the embedded context to improve query understanding. Godex can parse Godot scenes and scripts, identify functions, and map node connections and signals, but the scene parsing is still incomplete and unpolished. It often hallucinates or produces unnecessary output, and it doesn’t yet handle detailed analysis of node connections, signals, or wiring well.

I’m concerned about the token budget too—I’ve used about 18% of my Cursor subscription already, and bigger refactors or deeper development could exhaust it quickly. So I decided to test Godex for a week to see if it’s viable before deciding whether to keep iterating or pivot fully to game development.

The main bugs I’m still dealing with for Nucleate are inconsistent native executable behavior during install, Mac install failures on non-admin accounts with unclear errors, a broken in-app recorder on Mac, silent install failures without guidance, and general first-time install quality issues. For Godex, the biggest problems are hallucinations, incomplete scene parsing, lack of detailed node and signal analysis, and rapid token consumption.

I’m still figuring out how to present Nucleate’s core idea without sounding too polished or scripted, and how much effort to put into refining Godex’s output given the hallucination issues. I’m also wrestling with how to balance time between Godex development and actual game development, and whether it’s worth investing serious effort into marketing or just letting the product speak for itself. I’m curious which communities will engage with Godex—Reddit, Discord, or somewhere else—and whether the messaging tweaks for Nucleate will actually improve user perception. The next week’s test of Godex should help me decide if it’s worth continuing or if I need to pivot.

Two-Line Takeaway

Key step forward

The most important forward progress step is the release of Nucleate version one on Gumroad and the initial development and testing of the Godex AI assistant for game development.

Significant barrier or risk

The most important barrier or risk is the Mac installation failures on non-admin accounts for Nucleate and the hallucination issues, incomplete parsing, and rapid token consumption with Godex, which threaten usability and sustainability.

Keynotes

Decisions & Observations

• Decided to improve Mac install error messages for non-admin accounts instead of bypassing security restrictions.
• Adjusted Nucleate messaging to clarify typing speed claims after harsh feedback but uncertain if perception improved.

Issues & Next Steps

• Planning to try personal knowledge management subreddit’s self-promo space and targeting students with a free lecture transcription tool for organic engagement.
• Ongoing Godex issues: hallucinations, incomplete scene parsing, poor node/signal analysis, and rapid token consumption threatening subscription limits.
• Testing Godex for one week to evaluate viability before deciding to continue iteration or pivot fully to game development.
• Evaluating how to balance time between Godex development and actual game development, and whether to invest more in marketing or rely on product merit.
• Next step: Complete Godex one-week test to inform decision on further development or pivot.

Uncategorized Notes

• Released Nucleate version one on Gumroad; initial traction minimal despite promotion efforts on Discord and Notion.
• Developed Godex, an AI assistant for game development, with basic Qt app using Ollama’s API and embedded Godot docs for local RAG.
• Main Nucleate bugs to resolve: inconsistent native executable behavior, Mac install failures on non-admin accounts, broken in-app recorder on Mac, silent install failures, and first-time install quality issues.

Generated with Nucleate by Watchlight Studio — distilled from full transcription.