| 1 | # AI Workflow and Model Selection |
| 2 | |
| 3 | ## Model Selection Strategy |
| 4 | |
| 5 | **Haiku 4.5** (90% of Sonnet capability, 3x cost savings): |
| 6 | - Lightweight agents with frequent invocation |
| 7 | - Pair programming and code generation |
| 8 | - Worker agents in multi-agent systems |
| 9 | |
| 10 | **Sonnet 4.5** (Best coding model): |
| 11 | - Main development work |
| 12 | - Orchestrating multi-agent workflows |
| 13 | - Complex coding tasks |
| 14 | |
| 15 | **Opus 4.5** (Deepest reasoning): |
| 16 | - Complex architectural decisions |
| 17 | - Maximum reasoning requirements |
| 18 | - Research and analysis tasks |
| 19 | |
| 20 | ## Context Window Management |
| 21 | |
| 22 | Avoid last 20% of context window for: |
| 23 | - Large-scale refactoring |
| 24 | - Feature implementation spanning multiple files |
| 25 | - Debugging complex interactions |
| 26 | |
| 27 | Lower context sensitivity tasks: |
| 28 | - Single-file edits |
| 29 | - Independent utility creation |
| 30 | - Documentation updates |
| 31 | - Simple bug fixes |
| 32 | |
| 33 | ## Ultrathink + Plan Mode |
| 34 | |
| 35 | For complex tasks requiring deep reasoning: |
| 36 | 1. Use `ultrathink` for enhanced thinking |
| 37 | 2. Enable **Plan Mode** for structured approach |
| 38 | 3. "Rev the engine" with multiple critique rounds |
| 39 | 4. Use split role sub-agents for diverse analysis |
| 40 | |
| 41 | ## Build Troubleshooting |
| 42 | |
| 43 | If build fails: |
| 44 | 1. Use **build-error-resolver** agent |
| 45 | 2. Analyze error messages |
| 46 | 3. Fix incrementally |
| 47 | 4. Verify after each fix |
| 48 |