Add prompt-refactor meta-skill and SWE-Bench modular skills

[juanmichelini]

Overview

This PR introduces a new approach to managing prompt complexity by decomposing monolithic prompts into modular, reusable skills.

What's Added

1. prompt-refactor (Meta-Skill)

A meta-skill that helps decompose monolithic prompts into modular, reusable components:

• analyze_prompt.py script - Automated prompt analysis tool that identifies phases and structure
• skill-design-patterns.md - Reference documentation for best practices in skill design
• Workflow for analyzing, designing, and implementing modular skills

2. SWE-Bench Modular Skills (6 skills)

Six specialized skills created by refactoring the SWE-Bench prompt template:

• code-issue-analysis - Analyze bug reports and issue descriptions
• code-exploration - Explore codebases to understand structure and patterns
• test-reproduction - Create minimal reproduction scripts before fixing
• fix-analysis - Analyze problems and plan fixes with best practices
• code-implementation - Implement fixes following established patterns
• code-verification - Thoroughly test and verify implementations

Why This Matters

Before (Monolithic Prompt)

• 3874 characters, 63 lines of instructions in a single template
• Hard to maintain and update individual sections
• No reusability across different use cases
• Context window bloat with instructions that may not all be relevant

After (Modular Skills)

• 7 focused, reusable skills (average ~80 lines each)
• Each skill is independently maintainable
• Skills can be reused across different benchmarks (SWE-Bench, Commit0, etc.)
• Progressive disclosure - only loads relevant skills when needed
• Better context window management

Benefits

1. Reusability - Skills can be used across different benchmarks and workflows
2. Maintainability - Update one skill without affecting others
3. Testability - Each skill can be tested independently
4. Context Efficiency - Progressive disclosure loads only what's needed
5. Discoverability - Clear skill metadata makes it easy to understand when to use each skill

Example Usage

The prompt-refactor meta-skill can be used to refactor other monolithic prompts:

python skills/prompt-refactor/scripts/analyze_prompt.py path/to/prompt.j2

The SWE-Bench skills can be referenced in new prompt templates or used directly by OpenHands when solving coding issues.

Testing

All skills follow the established SKILL.md format with:

• ✅ Valid YAML frontmatter (name, description)
• ✅ Clear overview and usage guidance
• ✅ Step-by-step workflows
• ✅ Best practices and examples

Related Work

This is part of a larger effort to transform the SWE-Bench benchmark prompt architecture. A corresponding PR will be created for the benchmarks repository with a new skills.j2 template that references these modular skills.

Files Changed

• skills/prompt-refactor/SKILL.md - Meta-skill documentation
• skills/prompt-refactor/scripts/analyze_prompt.py - Prompt analysis tool
• skills/prompt-refactor/references/skill-design-patterns.md - Design patterns
• skills/code-issue-analysis/SKILL.md - Issue analysis skill
• skills/code-exploration/SKILL.md - Code exploration skill
• skills/test-reproduction/SKILL.md - Test reproduction skill
• skills/fix-analysis/SKILL.md - Fix analysis skill
• skills/code-implementation/SKILL.md - Implementation skill
• skills/code-verification/SKILL.md - Verification skill

Total: 1101 lines added across 9 files

Co-authored-by: openhands openhands@all-hands.dev

[openhands-ai]

Looks like there are a few issues preventing this PR from being merged!

• GitHub Actions are failing:
  • Check README.md in Skills

If you'd like me to help, just leave a comment, like

@OpenHands please fix the failing actions on PR #42 at branch `add-prompt-refactor-skill`

Feel free to include any additional details that might help me get this PR into a better state.

_{^{You can manage your notification settings}}