agent-output-formats（Agent 输出格式）

Agent Output Formats Skill

Standardized output formats for all agent types to ensure consistent communication and parsing across the autonomous development workflow.

When This Skill Activates

• Generating agent outputs
• Parsing agent responses
• Formatting research findings
• Creating planning documents
• Reporting implementation results
• Writing code reviews
• Keywords: "output", "format", "research", "planning", "implementation", "review"

Research Agent Output Format

Research agents (e.g., researcher, issue-creator, brownfield-analyzer) should structure outputs with these sections:

Template

## Patterns Found

[List of discovered patterns with examples]

- **Pattern Name**: Description
  - Example: Code snippet or reference
  - Use case: When to apply this pattern

## Best Practices

[Industry best practices and recommendations]

- **Practice Name**: Description
  - Benefit: Why this matters
  - Implementation: How to apply

## Security Considerations

[Security implications and requirements]

- **Security Concern**: Description
  - Risk: Potential vulnerabilities
  - Mitigation: How to address

## Recommendations

[Actionable recommendations for implementation]

1. **Recommendation**: Detailed guidance
   - Priority: High/Medium/Low
   - Effort: Time estimate
   - Impact: Expected benefit

Example Output

See examples/research-output-example.md for a complete example.

Planning Agent Output Format

Planning agents (e.g., planner, migration-planner, setup-wizard) should structure outputs with these sections:

Template

## Feature Summary

[Brief description of what will be built]

**Goal**: What this achieves
**Scope**: What's included/excluded
**Success Criteria**: How to measure success

## Architecture

[High-level design and component relationships]

**Components**: List of major components
**Data Flow**: How data moves through system
**Integration Points**: External dependencies

## Components

[Detailed component specifications]

### Component 1: [Name]
- **Purpose**: What it does
- **Responsibilities**: Core functions
- **Dependencies**: What it needs
- **Files**: Where it lives

## Implementation Plan

[Step-by-step implementation guide]

**Phase 1**: [Description]
1. Step one
2. Step two

**Phase 2**: [Description]
1. Step one
2. Step two

## Risks and Mitigations

[Potential issues and how to address them]

- **Risk**: Description
  - **Impact**: Severity and consequences
  - **Mitigation**: How to prevent or handle

Example Output

See examples/planning-output-example.md for a complete example.

Implementation Agent Output Format

Implementation agents (e.g., implementer, retrofit-executor) should structure outputs with these sections:

Template

## Changes Made

[Summary of what was implemented]

**Feature**: What was built
**Approach**: How it was implemented
**Design Decisions**: Key choices made

## Files Modified

[List of changed files with descriptions]

### Created Files
- `path/to/file.py`: Description of new file
- `path/to/test.py`: Test coverage

### Modified Files
- `path/to/existing.py`: Changes made
  - Added: New functionality
  - Modified: Updated behavior
  - Removed: Deprecated code

## Tests Updated

[Test coverage changes]

**New Tests**:
- Test file: What it covers
- Coverage: Percentage or lines

**Updated Tests**:
- Test file: What changed
- Reason: Why it was needed

## Next Steps

[Follow-up actions and recommendations]

1. **Action**: What needs to happen next
   - Owner: Who should do it
   - Priority: Urgency level
   - Blockers: Any dependencies

Example Output

See examples/implementation-output-example.md for a complete example.

Review Agent Output Format

Review agents (e.g., reviewer, security-auditor, quality-validator) should structure outputs with these sections:

Template

## Findings

[Overview of review results]

**Reviewed**: What was examined
**Scope**: What was checked
**Summary**: High-level results

## Code Quality

[Code quality assessment]

### Strengths
- **Aspect**: What's done well
  - Evidence: Specific examples

### Areas for Improvement
- **Issue**: What needs work
  - Severity: Critical/Major/Minor
  - Recommendation: How to fix
  - Location: Where the issue is

## Security

[Security analysis]

### Security Strengths
- **Protection**: What's secure
  - Implementation: How it's done

### Security Concerns
- **Vulnerability**: Potential issue
  - CWE Reference: Standard classification
  - Risk Level: High/Medium/Low
  - Remediation: How to fix

## Documentation

[Documentation assessment]

### Documentation Completeness
- **Aspect**: What's documented
  - Quality: How well it's done

### Documentation Gaps
- **Missing**: What needs docs
  - Priority: How important
  - Suggestion: What to add

## Verdict

[Final recommendation]

**Status**: ✅ APPROVED / ⚠️ APPROVED WITH CHANGES / ❌ NEEDS REVISION

**Rationale**: Why this verdict
**Blockers**: Must-fix issues (if any)
**Suggestions**: Nice-to-have improvements

Example Output

See examples/review-output-example.md for a complete example.

Commit Message Format

Commit message generator agents should follow conventional commits:

Template

<type>(<scope>): <subject>

<body>

<footer>

Types

• feat: New feature
• fix: Bug fix
• docs: Documentation only
• style: Formatting, no code change
• refactor: Code restructuring
• test: Adding tests
• chore: Maintenance tasks

Example

feat(skills): add agent-output-formats skill for standardized outputs

Extracts duplicated output format specifications from 15 agent prompts
into a reusable skill package following progressive disclosure architecture.

Token savings: ~3,000 tokens (200 tokens per agent × 15 agents)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Pull Request Format

PR description generator agents should follow this structure:

Template

## Summary

[Brief description of changes]

- Key change 1
- Key change 2
- Key change 3

## Test Plan

- [ ] Unit tests pass
- [ ] Integration tests pass
- [ ] Manual testing completed
- [ ] Documentation updated

## Related Issues

Closes #XXX

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Usage Guidelines

For Agent Authors

When creating or updating agent prompts:

1. Reference this skill in the "Relevant Skills" section
2. Remove duplicate format specifications from agent prompts
3. Trust progressive disclosure - full content loads when needed
4. Use consistent terminology from this skill

For Claude

When executing agents:

1. Load this skill when keywords match ("output", "format", etc.)
2. Follow format templates for structured outputs
3. Include all required sections for agent type
4. Maintain consistency across similar agents

Token Savings

By centralizing output formats in this skill:

• Before: ~250 tokens per agent for format specification
• After: ~50 tokens for skill reference
• Savings: ~200 tokens per agent
• Total: ~3,000 tokens across 15 agents (8-12% reduction)

Progressive Disclosure

This skill uses Claude Code 2.0+ progressive disclosure architecture:

• Metadata (frontmatter): Always loaded (~150 tokens)
• Full content: Loaded only when keywords match
• Result: Efficient context usage, scales to 100+ skills

When you use terms like "output format", "research findings", "planning document", or "code review", Claude Code automatically loads the full skill content to provide detailed guidance.

Examples

Complete example outputs are available in the examples/ directory:

• research-output-example.md: Sample research agent output
• planning-output-example.md: Sample planning agent output
• implementation-output-example.md: Sample implementation agent output
• review-output-example.md: Sample review agent output

Refer to these examples when generating agent outputs to ensure consistency and completeness.