Skip to main content

Learning System

Overview​

The Learning System tracks mistakes, preferences, and successes to build pattern recognition for future improvement. It extends the original mistake-tracking system with ACE (Agentic Context Engineering) skillbook functionality.

Components​

Learning Store​

The Learning Store persists learning entries and skills in .radium/_internals/learning/learning-log.json:

  • Mistakes: Errors made and corrected, with solutions
  • Preferences: User preferences and constraints
  • Successes: Successful patterns and approaches
  • Skills: Strategies organized by section in the skillbook

Skillbook​

The skillbook contains learned strategies organized by section:

  • task_guidance: How to approach and break down tasks
  • tool_usage: Best practices for using tools
  • error_handling: Error handling patterns
  • code_patterns: Code structure and patterns
  • communication: Communication strategies
  • general: General strategies

Each skill tracks:

  • Helpful count: Times this skill was helpful
  • Harmful count: Times this skill was harmful
  • Neutral count: Times this skill was neutral

CLI Commands​

List Learning Entries​

View all learning entries:

# List all entries
rad learning list

# Filter by category
rad learning list --category "Complex Solution Bias"

# JSON output
rad learning list --json

Add Mistake​

Record a mistake for future learning:

rad learning add-mistake \
  --category "Feature Creep" \
  --description "Added unnecessary features beyond requirements" \
  --solution "Stick to core requirements and avoid scope expansion"

Add Skill​

Add a skill to the skillbook:

rad learning add-skill \
  --section "task_guidance" \
  --content "Break complex tasks into smaller, manageable steps"

Tag Skill​

Tag a skill as helpful, harmful, or neutral:

# Tag as helpful
rad learning tag-skill --skill-id "skill-00001" --tag "helpful"

# Tag as harmful
rad learning tag-skill --skill-id "skill-00001" --tag "harmful" --increment 1

# Tag as neutral
rad learning tag-skill --skill-id "skill-00001" --tag "neutral"

Show Skillbook​

Display skills from the skillbook:

# Show all sections
rad learning show-skillbook

# Show specific section
rad learning show-skillbook --section "task_guidance"

# JSON output
rad learning show-skillbook --json

Standard Categories​

Mistakes are organized into standard categories:

  • Complex Solution Bias: Over-engineering solutions
  • Feature Creep: Adding unnecessary features
  • Premature Implementation: Jumping to code too quickly
  • Misalignment: Wrong direction or misunderstanding
  • Overtooling: Using too many tools unnecessarily
  • Preference: User preferences
  • Success: Successful patterns
  • Other: Uncategorized entries

Categories are automatically normalized (e.g., "complex" → "Complex Solution Bias").

Duplicate Detection​

The learning system prevents duplicate entries by detecting similar mistakes (60%+ word overlap). This prevents the skillbook from being cluttered with redundant information.

Integration with Vibe Check​

The learning system automatically integrates with Vibe Check:

  1. Mistake Logging: When oversight detects a mistake, it's logged to the learning store
  2. Pattern Extraction: Helpful and harmful patterns are extracted from oversight feedback
  3. Skillbook Updates: Patterns are converted to skills and added to the skillbook
  4. Context Injection: Learning context is injected into future oversight requests

Examples​

Example 1: Adding a Mistake​

$ rad learning add-mistake \
  --category "Complex Solution Bias" \
  --description "Created overly complex solution with unnecessary abstractions" \
  --solution "Simplify by removing unnecessary layers and using direct approach"

Mistake added successfully
  Category: Complex Solution Bias
  Description: Created overly complex solution with unnecessary abstractions
  Solution: Simplify by removing unnecessary layers and using direct approach

Example 2: Viewing Skills​

$ rad learning show-skillbook --section "task_guidance"

Skillbook

  task_guidance
    • [skill-00001]
      Break complex tasks into smaller, manageable steps
      Stats: Helpful: 5 | Harmful: 0 | Neutral: 1

    • [skill-00002]
      Start with core requirements before adding features
      Stats: Helpful: 3 | Harmful: 0 | Neutral: 0

Example 3: Tagging a Skill​

$ rad learning tag-skill --skill-id "skill-00001" --tag "helpful"

Skill skill-00001 tagged as helpful (increment: 1)

Best Practices​

  1. Be specific: Provide clear descriptions and solutions for mistakes
  2. Use standard categories: Stick to standard categories for better organization
  3. Tag skills regularly: Tag skills as helpful/harmful based on outcomes
  4. Review skillbook: Periodically review the skillbook to identify useful patterns
  5. Avoid duplicates: The system detects duplicates, but be mindful when adding entries

Troubleshooting​

Learning entries not persisting​

  • Check that .radium/_internals/learning/ directory exists
  • Verify file permissions for learning-log.json
  • Ensure workspace is properly initialized (rad init)

Skills not appearing in skillbook​

  • Verify the section name matches a standard section
  • Check that skills haven't been soft-deleted (use --include-invalid if needed)
  • Ensure the skillbook context is being generated correctly

Duplicate detection too strict/loose​

  • The 60% word overlap threshold is fixed
  • Use more specific descriptions to avoid false duplicates
  • Use less specific descriptions if legitimate entries are being rejected

References​