Version 3.2 of cline has been released, featuring the Plan & Act Modes which effectively distinguish between AI code thinking and action, highlighting the capabilities of cline’s ‘System Two’.

Planning Mode

• Optimized for situational collection and strategy

• Collaborates with you to formulate plans; does not write code

• Complete document reading capability for project understanding

• Focuses on breaking down complex requirements and formulating plans

Act Mode (the cline you are familiar with)

• Simplifies based on the provided context for ease of implementation

• Executes according to the established plan

• Maintains full access to cline’s build capabilities

cline responds that the benefits of incorporating Plan & Act Modes are obvious

1. Saves time in the long run

2. Produces more maintainable code

3. Captures edge cases early

4. Creates better documentation naturally

5. Makes iteration easier

Meanwhile, with the release of deepseek R1, cline has also begun to officially recommend this model.

A new workflow has been proposed, using deepseek R1 for planning and 3.5-Sonnet for code implementation.

• Planning with R1: $0.55/M tokens

• Planning with 3.5-Sonnet: ~$16/M tokens

• Cost reduction of 97% in the planning phase

However, this is not just about cost savings. As Austin Stolls pointed out in a recent analysis, R1 performs on par with or surpasses other models in reasoning benchmarks. When you combine this with the implementation capabilities of Sonnet, you get the best of both worlds.

Planning Phase (R1)

• In-depth architectural thinking

• Edge case exploration

• System design decisions

• Economical reasoning

Implementation Phase (3.5-Sonnet)

• Precise code generation

• Complex refactoring

• Bug fixing

• Battle-tested reliability

How to Use This Workflow

Want to try this approach? Here’s the process:

1. Start your task in planning mode using R1

2. Engage in a proper planning dialogue

3. When ready to write code, switch to Sonnet

4. Need to revise the plan? Switch back to R1

From experience, while this solution saves money, switching back and forth between models can be a bit overwhelming; hopefully, the next iteration will be more seamless.

References:

1. Everyone’s Talking About R1 vs O1 Benchmarks. But Here’s What Really Matters.

That concludes today’s content. The world of AI changes every day, and great tools continue to emerge. If you want to learn AI knowledge, master AI skills, and gain relevant careers, you can follow this public account. You can also add me on WeChat to join the AI enthusiast group chat.

In 2025, let’s progress together.

AI applications worth over 10 million + Feishu multi-dimensional table templates, surpassing AI courses on the market, free giveaway.

Strongly recommend the generative AI course (free learning with certification).

Custom solutions for agent workflows are available; feel free to consult.

Keywords: AI

Creating is not easy. Since you’ve read this far, if you find it useful, please subscribe to the public account, then give a thumbs up, see it, and share it.