Cline Best-of-Breed: Planning with DeepSeek R1, Code Implementation with 3.5-Sonnet

exist Cline On the same day that the Plan & Act model was released, DeepSeek introduced the R1 model. An interesting phenomenon ensued: developers spontaneously started using R1 for the planning phase and 3.5-Sonnet for code implementation. This was not prearranged, but based on a natural understanding of the advantages of both models.

New workflow for developers

The best practice that developers have found is:

1. Planning phase: Use the R1 model ($0.55/M tokens) to start a new task in Plan mode.
2. In-depth discussion: Have in-depth planning conversations to discuss architecture, edge cases, and possible solutions.
3. Code Implementation: When planning is complete, switch to the 3.5-Sonnet model for coding.
4. Iterative optimization: If you need to revisit the plan or encounter complex problems, switch back to the R1 model.

Double the cost and performance

The advantage of this workflow is not only in performance, but also in cost savings:

• R1 Model Planning: The cost is approximately $0.55/M tokens.
• 3.5-Sonnet model planning: The cost is about $16/M tokens.
• Cost reductions at the planning stage: Up to 97%.

In addition, as Austin Starks points out in his analysis, the R1 model is equal to or better than other models in terms of reasoning power. Combined with the proven ability of the 3.5-Sonnet model to be implemented in code, developers can achieve optimal results.

The two models complement each other's strengths

The key to the success of this model is that the characteristics of each model are fully utilized:

R1 model: ideal for the planning phase

• Deep Architecture Thinking: Ability to perform in-depth system architecture design.
• Edge case exploration: Good at identifying and dealing with potential edge cases.
• System design decisions: Ability to make sound system design decisions.
• Cost-effective reasoning: Cost advantage in reasoning tasks.

3.5 The Sonnet Model: Reliable Assurance of Code Enforcement

• Precise code generation: Ability to generate accurate and reliable code.
• Complex reconfiguration capabilities: Complex code refactoring is possible.
• Quick bug fix: Ability to quickly locate and fix errors in code.
• Proven reliability: Highly reliable in practical applications.

How to apply this workflow

1. Plan mode is activated: Use the R1 model to start Plan mode in Cline.
2. Planning for Dialogue: Have a full planning conversation to determine program details.
3. Switch to Sonnet: After completing the planning, switch to the 3.5-Sonnet model to write the code.
4. Iterative optimization: If you need to modify the plan, switch back to the R1 model to make adjustments.

summarize

It's not just about R1 vs. o1 or 3.5-Sonnet, it's about using the right tool at the right stage. the future of AI development isn't about finding a one-size-fits-all model, it's about being flexible enough to utilize the strengths of different models based on the needs of the task.

It's worth noting that this best practice was not pre-designed, but rather spontaneously explored by Cline developers in real-world use. This discovery reflects the wisdom of the community, and the results are a testament to the value of this workflow.

Have you tried this R1/Sonnet workflow? What other model combinations have you found to be effective?

Our goal is to build better software by taking advantage of different models and using the most appropriate tools at each stage of software development.