---

name: ralph-wiggum description: Self-critique loop that catches mistakes before finalizing - "I'm in danger" detection tools:

• Read
• Grep
• Bash

---

Ralph Wiggum Self-Critique Loop

Named after the Simpsons character who says "I'm in danger" - this skill helps you recognize when you're about to make a mistake.

When This Activates

Automatically engage self-critique when:

• You just wrote or modified code
• You're about to run a destructive command
• You made multiple changes in sequence
• The task seemed "too easy"
• You're uncertain about something

The Ralph Loop

After completing any significant action, pause and ask:

1. Reality Check (Am I in Danger?)

Copy□ Did I actually solve the problem, or just change something?
□ Did I read the error message carefully?
□ Am I making assumptions I haven't verified?
□ Does this "fix" match what I claimed I would do?

2. Obvious Mistakes Check

Copy□ Syntax errors (missing brackets, typos)
□ Wrong file edited
□ Deleted something important
□ Hardcoded values that should be variables
□ Left debug code/print statements
□ Broke something that was working

3. Edge Case Scan

Copy□ What if input is empty/null?
□ What if input is huge?
□ What if this runs twice?
□ What if network fails?
□ What happens on error?

4. Confidence Score

Rate yourself 1-10:

• 1-3: I'm definitely wrong, let me start over
• 4-6: Something feels off, let me verify
• 7-8: Probably right, quick sanity check
• 9-10: Confident, but verify anyway

5. Verification Action

Based on score, take action:

Score 1-6: Full review

Copy# Run linter
ruff check path/to/file.py

# Run type checker
mypy path/to/file.py

# Run tests
pytest -x path/to/test_file.py

# Diff what changed
git diff path/to/file.py

Score 7-8: Quick check

Copy# Just the linter
ruff check path/to/file.py --select=E,F

# Or syntax check
python -m py_compile path/to/file.py

Score 9-10: Minimal check

• Re-read the code once
• Ensure it matches the task

Common "I'm in Danger" Signals

Red Flags

• "This should work..." (but you haven't tested)
• Making the same fix for the third time
• The fix is longer than the original code
• You're not sure what the code does
• Copying code you don't understand

Yellow Flags

• Task took much less time than expected
• You skipped reading the whole function
• You assumed the test coverage is good
• You're tired or distracted

Example Ralph Loop

CopyUser: Fix the bug in calculate_total()

[After making change]

🔍 Ralph Check:
1. Reality: Did I fix the actual bug? Let me re-read the issue...
   - Issue was: negative numbers cause overflow
   - My fix: added abs() call
   - ⚠️ Wait, that changes the behavior for valid negatives!

2. Confidence: 4/10 - Something's wrong

3. Action: Let me reconsider...
   - The real fix should validate input, not silently change it
   - Adding: if value < 0: raise ValueError("...")

4. Re-check: Now confidence 8/10

5. Verify: Run tests... ✓ Pass

Trigger Phrases

Say these to invoke Ralph mode:

• "ralph check" - Run full self-critique on last action
• "am I in danger?" - Quick reality check
• "let me verify" - Run verification commands
• "confidence check" - Rate and explain confidence level

Integration with Adversarial Pipeline

Ralph is your internal critic. For external validation:

Copy1. You make changes
2. Ralph checks your work (self-critique)
3. If Ralph passes → Send to Critic agent (external review)
4. If Critic finds issues → Fixer agent repairs
5. Verifier confirms

Ralph catches ~60% of issues before they reach the critic, saving cycles.