Best Practices

Follow these guidelines to get the best results from Open SWE.

Prompting Tips

Be Clear and Direct

  • Include specific file paths and function names in your requests. Open SWE preforms best when its given a clear starting point.
  • Provide concrete examples of what you want to achieve. Describe the end state you want to reach so Open SWE knows what it’s working towards.

Create Custom Rules

Create an AGENTS.md file in your repository root to provide project-specific context. This helps Open SWE understand your codebase conventions and requirements.
See the Custom Rules page for detailed guidance on setting up your AGENTS.md file.

Keep Tasks Well-Scoped

  • Focus on one specific feature or fix per request
  • Break large changes into smaller, manageable tasks
  • Avoid combining multiple unrelated changes in a single request

Avoid Multiple Tasks

Submit separate requests for different features or fixes. This allows Open SWE to:
  • Generate more focused plans
  • Provide better error handling
  • Make changes easier to review

Model Selection

  • Claude Sonnet 4 (Default): The default model for planning, writing code, and reviewing changes. This model offers the best balance of performance, speed and cost.
  • Claude Opus 4: A larger, more powerful model for difficult, or open-ended tasks. Opus 4 is more expensive and slower, but will provide better results for complex tasks.

Avoid Other Models

Although Open SWE allows you to select any model from Anthropic, OpenAI and Google, its prompts are tuned specifically for Anthropic models, and other providers will not preform as well.

Mode Selection

open-swe vs open-swe-max

open-swe: Uses Claude Sonnet 4
  • Suitable for most development tasks
  • Faster execution
  • Cost-effective
open-swe-max: Uses Claude Opus 4 (only for the planning and code writing agents)
  • For complex tasks requiring advanced reasoning
  • Higher quality output for challenging problems
  • More expensive but better for difficult tasks

Auto vs Manual Labels

Auto Mode (-auto labels) It’s recommended to use the auto mode for most tasks. Open SWE is very good at planning, and in most cases it does not need a manual review before execution. If you’re running Open SWE against an open-ended or very complex task, you may want to use manual mode to review the plan before execution.

Label Reference

  • open-swe: Manual mode with Sonnet 4
  • open-swe-auto: Auto mode with Sonnet 4
  • open-swe-max: Manual mode with Opus 4
  • open-swe-max-auto: Auto mode with Opus 4
In development environments, append -dev to all labels (e.g., open-swe-dev, open-swe-auto-dev).