Monday, June 1, 2026June 1, 2026
- Ran the Godot project and checked the map generation results
- Today’s actual project work centered on running Godot and viewing the map state
- The model could not reliably interpret images and screen results, making map improvement decisions difficult
- Custom web‑search MCP in a private repository
- Adjusted
web-search-mcp to match the desired search flow and output
- Found the usability of web‑search MCP in tools other than Qwen Code to be poorer than expected, leading to frustration
- Qwen Code, Codex, OpenCode comparative tests
- Qwen Code frequently repeated answers and retried planning
- Codex and OpenCode showed limitations in understanding Godot context and image interpretation
- Some tools failed while planning and sub‑agents for research queried the filesystem, causing flow mismatches
- Godot‑specific model experiments
- Used Hugging Face
bekoozkan/godot-gemma-4-e4b-it-GGUF model on RunPod
- Reference: https://huggingface.co/bekoozkan/godot-gemma-4-e4b-it-GGUF
- Several models tended to suggest code patterns from older Godot versions rather than the latest, prompting a trial of a specialized model
- However, this specialized model had few parameters and weak context understanding, so it was not as stable as hoped
- AI tool usage reflection
- Qwen Code seems unlikely to overcome repetitive issues without proper tuning, regardless of the model used
- Godot work is hard to consider complete without verification on the actual execution screen
- I feel this field may not be a good fit for me
- Starting tomorrow, I plan to collect the necessary Godot work patterns myself and attempt fine‑tuning
- Development retrospective: docs/retrospectives/2026-06-01.md
- Meeting minutes: docs/meetings/2026-06-01.md
May 2026