1.5 KiB
1.5 KiB
LLM Strategy & Workflow
Fallback Strategy (After 3 attempts / 10 minutes):
When initial attempts fail, sequentially try these LLMs:
- DeepSeek-V3-0324 (good architect, adventure, reliable, let little errors just to be fixed by gpt-*)
- gpt-5-chat (slower, let warnings...)
- gpt-oss-120b
- Claude (Web): Copy only the problem statement and create unit tests. Create/extend UI.
Development Workflow:
- One requirement at a time with sequential commits
- On unresolved error: Stop and use add-req.sh, and consult Claude for guidance. with DeepThining in DeepSeek also, with Web turned on.
- Change progression: Start with DeepSeek, conclude with gpt-oss-120b
- If a big req. fail, specify a @code file that has similar pattern or sample from official docs.
- Warning removal: Last task before commiting, create a task list of warning removal and work with cargo check.
- Final validation: Use prompt "cargo check" with gpt-oss-120b
- Be humble, one requirement, one commit. But sometimes, freedom of caos is welcome - when no deadlines are set.
- Fix manually in case of dangerous trouble.
- Keep in the source codebase only deployed and tested source, no lab source code in main project. At least, use optional features to introduce new behaviour gradually in PRODUCTION.
- Transform good articles into prompts for the coder.
- Switch to libraries that have LLM affinity (LLM knows the library, was well trained).
- Ensure 'continue' on LLMs, they can EOF and say are done, but got more to output.