Autonomous Iteration Mode

Enable self-referential iteration to complete complex tasks that require multiple attempts, trial-and-error exploration, or persistent effort until success.

The Ralph Wiggum Methodology

This skill implements key insights from the Ralph Wiggum methodology:

1. Stop Hook Iteration: Don't exit until success criteria are met
2. Completion Promises: Explicit signals when task is complete
3. Failures as Data: Each failure informs the next attempt
4. Deterministic Retry: Systematic approach to retrying

Process

1. Define Success Criteria

Before iterating, establish clear success criteria:

## Iteration Goal
[What are we trying to achieve?]

## Success Criteria
- [ ] Criterion 1 (e.g., "tests pass")
- [ ] Criterion 2 (e.g., "no type errors")
- [ ] Criterion 3 (e.g., "feature works as expected")

## Max Attempts
[Number, typically 5-10]

Write this to:

~/.claude-mind/state/iteration-goal.md

2. Create Iteration State

Track progress in:

~/.claude-mind/state/iteration-state.json

{
  "goal": "Description of goal",
  "criteria": ["criterion 1", "criterion 2"],
  "maxAttempts": 5,
  "currentAttempt": 0,
  "status": "in_progress",
  "attempts": [],
  "startedAt": "ISO8601",
  "updatedAt": "ISO8601"
}

3. Iteration Loop

For each attempt:

1. Increment attempt counter

2. Execute the task

3. Evaluate against criteria

4. Record outcome:

   • What was tried
   • What happened
   • What was learned
   • What to try next (if not successful)

5. Decide next action:

   • If ALL criteria met → Mark SUCCESS, exit
   • If max attempts reached → Mark FAILED, summarize learnings, exit
   • Otherwise → Continue to next attempt with adjusted approach

4. Attempt Record Format

Add to the

attempts

array:

{
  "number": 1,
  "action": "What was done",
  "outcome": "What happened",
  "criteria_met": ["criterion 1"],
  "criteria_failed": ["criterion 2"],
  "learning": "What this taught us",
  "next_approach": "What to try differently"
}

5. Completion Signals

When complete, update state:

{
  "status": "success" | "failed" | "abandoned",
  "completedAt": "ISO8601",
  "summary": "Final outcome description",
  "total_attempts": 3,
  "key_learnings": ["learning 1", "learning 2"]
}

Commands

Start Iteration

# Initialize new iteration
~/.claude-mind/system/bin/iterate-start "goal description" --max-attempts 5 --criteria "tests pass" --criteria "builds clean"

Check Status

# View current iteration state
~/.claude-mind/system/bin/iterate-status

Record Attempt

# Record an attempt outcome
~/.claude-mind/system/bin/iterate-record --success | --failure --action "what was done" --learning "what was learned"

Complete Iteration

# Mark iteration complete
~/.claude-mind/system/bin/iterate-complete --success | --failed "summary"

Guidelines

• Be systematic: Don't repeat the exact same approach
• Learn from failures: Each failure should inform the next attempt
• Stay focused: Don't drift from the original goal
• Know when to stop: Max attempts prevents infinite loops
• Document everything: Future iterations benefit from past learnings

Integration with Stop Hook

The iteration system works with Claude Code's Stop hook. When iteration is active:

1. Stop hook checks

   ~/.claude-mind/state/iteration-state.json

2. If status is

   in_progress

   and criteria not met:
   • Hook provides feedback: "Iteration in progress, X criteria remaining"
   • Suggests next action based on last attempt
3. This creates a self-referential loop until success or max attempts

Example Usage

User: I need you to fix the build errors, keep trying until it compiles

Claude: I'll enter iteration mode for this task.

[Creates iteration state with goal "fix build errors" and criterion "build succeeds"]

Attempt 1: Fixed missing import
- Outcome: 3 errors remain
- Learning: Type errors in DataProcessor.swift

Attempt 2: Fixed type mismatches
- Outcome: 1 error remains
- Learning: Optional unwrapping issue

Attempt 3: Added nil check
- Outcome: Build succeeds!
- All criteria met

[Marks iteration SUCCESS]

The build now compiles cleanly after 3 attempts.

Anti-Patterns

• Don't iterate without clear success criteria
• Don't continue iterating if fundamentally stuck (ask for help instead)
• Don't use for tasks that need human decision-making
• Don't exceed max attempts without human approval