majiayu000 Conversation Improvement Skill
Quick-trigger skill for analyzing conversations and improving AI behavior. This skill provides a streamlined version of the conversation-improver agent for immediate feedback and fixes.
When This Activates
This skill auto-activates when you:
- Say "improve", "improve chat", or "/improve"
- Ask about what went wrong
- Want to analyze conversation issues
- Need to fix AI behavior problems
- Ask for retrospective analysis
Quick Improvement Process
1. Identify the Issue (Ask if unclear)
Quick questions:
- What specific action didn't match your expectations?
- What should have happened instead?
- Is this recurring or one-time?
- Should the fix apply broadly or in specific contexts?
2. Analyze Root Cause
Quickly identify:
- Scope Creep - Did extra work
- Wrong Interpretation - Misunderstood request
- Missing Guardrail - No rule to prevent this
- Over-Caution - Too many questions
- Under-Caution - Acted without confirmation
- Tool Misuse - Wrong tool choice
3. Propose Fix
Target the right location:
- Skill → Domain-specific behavior (
.github/skills/) - Custom Instructions → Project-wide rules (
.github/copilot-instructions.md) - Agent → Workflow-specific (
.github/agents/)
4. Present Solution
## Improvement Analysis
**Issue**: [What went wrong]
**Root Cause**: [Why it happened]
**Proposed Fix**: [Where and what to change]
### Target File
[Path to file that needs updating]
### Change
[Exact text to add/replace]
### Risk
Low/Medium/High - [Why]
Quality Checks
Before proposing changes, verify:
- ✓ Specific enough to prevent the issue?
- ✓ General enough for future cases?
- ✓ No conflicts with existing rules?
- ✓ Testable - can user verify?
- ✓ Minimal - smallest change?
Output Options
Based on user preference:
- Analysis Only - Report without changes
- Proposed Patch - Show changes for approval
- Apply Changes - Make changes and summarize
Anti-Patterns to Avoid
- ❌ Shotgun fixes (multiple files for one issue)
- ❌ Over-specification (too narrow rules)
- ❌ Vague language ("be careful", "consider")
- ❌ Rule duplication (same rule in multiple places)
Example Usage
User: "improve - you kept refactoring files I didn't ask you to touch"
Response:
### Issue: Scope Creep
AI modified files outside the requested scope.
### Root Cause
Instructions allow "fixing issues in vicinity" but "vicinity" is subjective.
### Proposed Fix
**Target**: `.github/copilot-instructions.md`
Add this rule:
> Only modify files explicitly mentioned in the request.
> Fix critical issues ONLY if they cause build/test failures.
> "Vicinity" means same function—never other files.
**Risk**: Low - Narrows behavior without blocking legitimate fixes.
Would you like me to apply this change?
For More Complex Analysis
For deeper investigation or complex issues, invoke the full conversation-improver agent:
/agent
> Select: conversation-improver
This provides comprehensive root cause analysis with full validation checks.
Remember
- Focus on patterns, not individual mistakes
- Prefer minimal, surgical changes
- Get user confirmation before applying
- Document why the change helps