AugmentClaude
BackendPM

Cross-Task Learner

By jmagly· jmagly/aiwg· 0

Learn patterns from past tasks and share them across multiple agent loops.

Installation

  1. Make sure Claude is on your device and in your terminal.

    Skills load from ~/.claude/skills/ when Claude Code starts up — so you need it on your machine first. If you don't have it yet, install it once with the command below, then run claude in any terminal to verify.

    One-time setup
    npm i -g @anthropic-ai/claude-code

    Already have it? Skip ahead.

  2. Paste into Claude Code or into your terminal.
    Install
    Sign up to copy
    Sign up free to reveal

    This copies the whole skill folder into ~/.claude/skills/cross-task-learner-jmagly/ — the SKILL.md plus any scripts, reference docs, or templates the skill ships with. Safe default: works for every skill.

    Faster alternative (instruction-only skills)

    Skips the clone and grabs only the SKILL.md file. Don't use this if the skill ships Python scripts, reference markdowns, or asset templates — they won't be downloaded and the skill will fail when it tries to load them.

    Quick install (SKILL.md only)
    Sign up to copy
    Sign up free to reveal
  3. Restart Claude Code.

    Quit and reopen Claude Code (or any other agent that loads from ~/.claude/skills/). New skills are picked up on startup.

  4. Just ask Claude.

    Skills auto-activate when your request matches the skill's description — no slash command needed. Trigger phrases live in the skill's own frontmatter; you can read them in the “What this skill does” section above.

Prefer to read the source first? Open on GitHub.

When Claude uses it

Enable agent loops to learn from similar past tasks and share patterns across loops

What this skill does

Cross-Task Learner Skill

Enable agent loops to learn from similar past tasks and share discovered patterns across multiple concurrent or sequential loops.

Research Foundation: REF-013 MetaGPT - 159% improvement with shared state

Version 2.0: Multi-loop awareness with loop_id tracking

Overview

This skill provides two core capabilities:

  1. Pattern Extraction - On loop completion, extract reusable patterns from execution history
  2. Pattern Injection - On loop start, inject relevant patterns from previous loops

Benefits

BenefitImpact
Faster resolutionPatterns eliminate redundant debugging
Higher success ratesProven approaches applied automatically
Accumulated wisdomSystem gets smarter over time
Anti-pattern detectionFailed approaches flagged and avoided

Research Basis

From REF-013 MetaGPT:

  • 159% improvement with shared state across agents
  • Publish-subscribe pattern enables knowledge sharing
  • Structured outputs become inputs for other agents
  • Memory persistence critical for cross-session learning

Pattern Extraction (On Loop Completion)

Trigger

  • Agent loop completion (success, partial, or failure)
  • Manual extraction request via aiwg ralph-extract-patterns {loop_id}

Process

extraction_steps:
  1_analyze_loop_history:
    - Load loop state from .aiwg/ralph/loops/{loop_id}/state.json
    - Load iteration analytics
    - Load debug memory
    - Load reflection history

  2_identify_error_fix_pairs:
    - Scan iterations for test failures
    - Identify fixes that resolved errors
    - Extract error signature + fix approach
    - Compute initial success rate (1.0 for first occurrence)

  3_identify_successful_approaches:
    - Analyze task category (testing, debugging, refactoring, etc.)
    - Extract step sequence that led to success
    - Note tools used and iteration count
    - Identify preconditions and benefits

  4_identify_failure_patterns:
    - Detect repeated same errors (anti-patterns)
    - Note approaches that led to scope creep
    - Flag patterns that caused quality degradation
    - Record better alternatives if discovered

  5_extract_code_templates:
    - Identify successful code changes
    - Generalize with placeholders
    - Document use case and placeholders
    - Tag by language and purpose

  6_check_for_duplicates:
    - Compare against existing patterns in registry
    - Merge if >80% similar
    - Update usage count and success rate if duplicate

  7_store_in_registry:
    - Add to .aiwg/ralph/shared/patterns/{type}-patterns.json
    - Update patterns index for semantic search
    - Link to source loop_id

  8_update_effectiveness_metrics:
    - Increment pattern counts
    - Update cross-loop benefit statistics
    - Log extraction event

Example Extraction

Input (from loop ralph-fix-auth-a1b2c3d4):

iteration_2:
  error:
    type: "TypeError"
    message: "Cannot read property 'email' of null"
    location: "src/auth/validate.ts:42"

iteration_3:
  fix_applied:
    description: "Added null check for user object"
    diff: |
      + if (user == null) {
      +   throw new ValidationError("User is required");
      + }
  test_results:
    passed: 12
    failed: 0

Output (extracted pattern):

pattern_id: "pat-error-null-check-015"
type: "error_pattern"
error_signature:
  error_type: "TypeError"
  error_pattern: "Cannot read property '.*' of null"
  error_location_hints:
    - "*.ts:validate*"
fix_approach:
  description: "Add null check before property access"
  fix_category: "add_null_check"
  code_template: |
    if ({{variable}} == null) {
      throw new ValidationError("{{message}}");
    }
  code_template_language: "typescript"
source_loops:
  - loop_id: "ralph-fix-auth-a1b2c3d4"
    timestamp: "2026-02-02T15:00:00Z"
    contributed_by: "software-implementer"
success_rate: 1.0
usage_count: 1
first_discovered: "2026-02-02T15:00:00Z"
last_used: "2026-02-02T15:00:00Z"
tags:
  - "typescript"
  - "null-safety"
  - "validation"

Configuration

# In aiwg.yml or .aiwg/config.yml
ralph:
  cross_loop_learning:
    extraction:
      enabled: true
      auto_extract_on_completion: true
      min_success_rate_threshold: 0.6
      min_usage_count_for_evaluation: 3
      extract_code_templates: true
      merge_similar_patterns: true
      similarity_threshold: 0.80

Pattern Injection (On Loop Start)

Trigger

  • Agent loop start
  • Manual injection request via aiwg ralph-inject-patterns {loop_id}

Process

injection_steps:
  1_analyze_task_description:
    - Extract task text
    - Identify task category (testing, debugging, refactoring, etc.)
    - Generate task embedding for semantic matching

  2_search_error_patterns:
    - Query error patterns by task category
    - Match error signatures to likely error types
    - Filter by min success rate (default 0.6)
    - Sort by effectiveness

  3_search_success_patterns:
    - Query success patterns by task category match
    - Use semantic similarity on task description
    - Filter by min success rate
    - Sort by average iterations (lower is better)

  4_search_anti_patterns:
    - Query anti-patterns by task category
    - Identify failure modes to avoid
    - Include better alternatives

  5_search_code_templates:
    - Query templates by language and task type
    - Filter by success rate
    - Sort by usage count

  6_filter_and_rank:
    - Combine all pattern types
    - Remove duplicates
    - Rank by relevance × effectiveness
    - Take top-k (default k=5)

  7_inject_into_context:
    - Format patterns for display
    - Add to loop context as "Cross-Loop Learning Context"
    - Track which patterns were injected (for effectiveness measurement)

  8_track_usage:
    - Log pattern injection event
    - Record loop_id and injected pattern_ids
    - Enable later effectiveness analysis

Example Injection

Input (loop ralph-fix-validation-b2c3d4e5 starting):

task: "Fix TypeScript type errors in user validation"
category: "debugging"

Patterns Retrieved:

retrieved_patterns:
  error_patterns:
    - pattern_id: "pat-error-null-check-015"
      relevance: 0.95
      success_rate: 1.0
      usage_count: 1

    - pattern_id: "pat-error-type-mismatch-003"
      relevance: 0.88
      success_rate: 0.94
      usage_count: 18

  success_patterns:
    - pattern_id: "pat-success-type-fixing-002"
      relevance: 0.92
      success_rate: 0.87
      average_iterations: 2.8

  anti_patterns:
    - pattern_id: "pat-anti-premature-optimization-001"
      relevance: 0.65
      failure_rate: 0.85

Injected Context (top 5 patterns):

## Cross-Loop Learning Context

Patterns from previous loops that may help with this task:

### Error Patterns

1. **TypeError: null property access** (100% success, 1 use)
   - Error: "Cannot read property '.*' of null"
   - Fix: Add null check before property access
   - Template:
     ```typescript
     if ({{variable}} == null) {
       throw new ValidationError("{{message}}");
     }
     ```
   - Source: ralph-fix-auth-a1b2c3d4

2. **TypeError: type mismatch** (94% success, 18 uses)
   - Error: "Type 'X' is not assignable to type 'Y'"
   - Fix: Update interface definition or add type assertion
   - Source: 18 previous loops

### Success Patterns

1. **Type error fixing approach** (87% success, avg 2.8 iterations)
   - Steps:
     1. Run tsc --noEmit to see all errors
     2. Group errors by file/type
     3. Fix interface definitions first
     4. Fix usages second
     5. Verify with tsc again
   - Source: 12 successful loops

### Anti-Patterns to Avoid

1. **Premature optimization** (85% failure rate)
   - Don't optimize before tests pass
   - Complete primary task first
   - Better alternative: pat-success-refactor-module-002

Configuration

ralph:
  cross_loop_learning:
    injection:
      enabled: true
      top_k_patterns: 5
      min_success_rate: 0.6
      max_patterns_injected: 10
      include_error_patterns: true
      include_success_patterns: true
      include_anti_patterns: true
      include_code_templates: true

Multi-Loop Awareness (Version 2.0)

Loop ID Tracking

All patterns now track their source loop IDs:

pattern:
  pattern_id: "pat-error-null-check-015"
  source_loops:
    - loop_id: "ralph-fix-auth-a1b2c3d4"
      timestamp: "2026-02-02T15:00:00Z"
      contributed_by: "software-implementer"
    - loop_id: "ralph-fix-validation-b2c3d4e5"
      timestamp: "2026-02-02T16:30:00Z"
      contributed_by: "debugger"

Aggregation Across Loops

Patterns aggregate learnings from multiple loops:

Example: Pattern seen in 3 loops

pattern_id: "pat-error-type-mismatch-003"
source_loops: [ralph-a, ralph-b, ralph-c]
usage_count: 18  # Total across all loops
success_rate: 0.94  # Aggregated success rate (17/18)

effectiveness_trend:
  - timestamp: "2026-02-01T10:00:00Z"
    success_rate_snapshot: 1.0
    sample_size: 1
    source_loop: "ralph-a"
  - timestamp: "2026-02-01T14:00:00Z"
    success_rate_snapshot: 0.90
    sample_size: 10
    source_loops: ["ralph-a", "ralph-b"]
  - timestamp: "2026-02-02T15:00:00Z"
    success_rate_snapshot: 0.94
    sample_size: 18
    source_loops: ["ralph-a", "ralph-b", "ralph-c"]

Cross-Loop Pattern Updates

When a loop uses a pattern, it updates the shared registry:

on_pattern_application:
  1_increment_usage_count:
    - pattern.usage_count += 1

  2_update_success_rate:
    - If application succeeded:
        pattern.successful_applications += 1
    - Else:
        pattern.failed_applications += 1
    - pattern.success_rate = successful / (successful + failed)

  3_update_effectiveness_trend:
    - Add new snapshot with current success rate and sample size

  4_add_source_loop:
    - If loop_id not in source_loops:
        pattern.source_loops.append({
          loop_id: current_loop_id,
          timestamp: now,
          contributed_by: current_agent
        })

  5_update_last_used:
    - pattern.last_used = now

Pattern Lineage

Patterns maintain full lineage across loops:

Pattern: pat-error-null-check-015

Discovered: ralph-fix-auth-a1b2c3d4 (2026-02-02T15:00:00Z)
Applied:    ralph-fix-validation-b2c3d4e5 (2026-02-02T16:30:00Z) ✓
Applied:    ralph-fix-user-c3d4e5f6 (2026-02-02T18:00:00Z) ✓
Applied:    ralph-fix-api-d4e5f6a7 (2026-02-02T19:15:00Z) ✗

Success rate: 75% (3/4)

Effectiveness Measurement

Metrics Tracked

All metrics now support multi-loop aggregation:

effectiveness_metrics:
  total_patterns: 47
  patterns_by_type:
    error_patterns: 24
    success_patterns: 15
    anti_patterns: 5
    code_templates: 3

  pattern_usage_stats:
    total_applications: 156
    successful_applications: 142
    failed_applications: 14
    overall_success_rate: 0.91

  cross_loop_benefit:
    loops_with_pattern_injection: 23
    loops_without_pattern_injection: 8
    average_iterations_with: 3.2
    average_iterations_without: 5.1
    improvement_percentage: 37.3  # (5.1 - 3.2) / 5.1 * 100

  last_updated: "2026-02-02T20:00:00Z"

Per-Loop Impact

Each loop tracks which patterns it used:

loop_state:
  loop_id: "ralph-fix-validation-b2c3d4e5"
  cross_loop_learning:
    injected_patterns:
      - pattern_id: "pat-error-null-check-015"
        applied: true
        successful: true
        iteration_used: 2

      - pattern_id: "pat-error-type-mismatch-003"
        applied: true
        successful: true
        iteration_used: 4

    extracted_patterns:
      - pattern_id: "pat-error-optional-types-016"
        new: true
        success_rate: 1.0

    improvement_estimate:
      baseline_iterations: 6  # Estimated without patterns
      actual_iterations: 4
      improvement: 33.3%

Storage Structure

.aiwg/ralph/
├── shared/                          # Cross-loop shared state
│   ├── patterns/
│   │   ├── error-patterns.json      # Error → fix patterns
│   │   ├── success-patterns.json    # Success approaches
│   │   ├── anti-patterns.json       # Failure patterns
│   │   └── code-templates.json      # Reusable code
│   ├── archive/                     # Pruned patterns
│   │   ├── 2026-01-patterns.json
│   │   └── 2026-02-patterns.json
│   └── effectiveness-metrics.json   # Cross-loop metrics
│
├── memory/                          # Cross-task memory
│   ├── task-index.json              # Semantic task index
│   ├── embeddings/                  # Task embeddings
│   ├── reflections/                 # Reflexion-style reflections
│   └── patterns/                    # Patterns integrated here too
│
└── loops/                           # Per-loop state
    ├── ralph-fix-auth-a1b2c3d4/
    │   ├── state.json               # Includes pattern usage
    │   └── analytics/
    │       └── analytics.json       # Pattern effectiveness
    └── ralph-fix-validation-b2c3d4e5/
        └── ...

CLI Commands

Extract Patterns Manually

# Extract from completed loop
aiwg ralph-extract-patterns ralph-fix-auth-a1b2c3d4

# Extract from all completed loops
aiwg ralph-extract-patterns --all

# Extract specific pattern types only
aiwg ralph-extract-patterns ralph-fix-auth-a1b2c3d4 --types error,success

Inject Patterns Manually

# Inject patterns into running loop
aiwg ralph-inject-patterns ralph-current-loop-e5f6a7b8

# Inject with custom filters
aiwg ralph-inject-patterns ralph-current-loop-e5f6a7b8 \
  --min-success-rate 0.8 \
  --top-k 10

List Patterns

# List all patterns
aiwg ralph-patterns list

# Filter by type
aiwg ralph-patterns list --type error

# Filter by success rate
aiwg ralph-patterns list --min-success-rate 0.8

# Filter by source loop
aiwg ralph-patterns list --from-loop ralph-fix-auth-a1b2c3d4

Show Pattern Details

aiwg ralph-patterns show pat-error-null-check-015

Output:

Pattern: pat-error-null-check-015
Type: Error Pattern
Created: 2026-02-02T15:00:00Z
Last Used: 2026-02-02T19:15:00Z

Effectiveness:
  Success Rate: 75% (3/4 applications)
  Usage Count: 4
  Average Impact: -2.0 iterations

Source Loops:
  - ralph-fix-auth-a1b2c3d4 (discovered)
  - ralph-fix-validation-b2c3d4e5 (applied successfully)
  - ralph-fix-user-c3d4e5f6 (applied successfully)
  - ralph-fix-api-d4e5f6a7 (applied, failed)

Error Signature:
  Type: TypeError
  Pattern: "Cannot read property '.*' of null"

Fix Approach:
  Category: add_null_check
  Template: |
    if ({{variable}} == null) {
      throw new ValidationError("{{message}}");
    }

Tags: typescript, null-safety, validation

Prune Patterns

# Dry run
aiwg ralph-patterns prune --dry-run

# Prune low-success patterns
aiwg ralph-patterns prune --min-success-rate 0.5

# Prune old unused patterns
aiwg ralph-patterns prune --max-age-days 90

Integration with Other Skills

With Reflection Injection

Patterns complement reflections:

  • Reflections: What was learned in this specific loop
  • Patterns: Reusable knowledge across all loops

Both are injected into new loop context.

With Auto-Test Execution

Patterns inform test execution:

  • Error patterns suggest which tests to run
  • Success patterns guide test-driven approaches
  • Code templates provide test setup patterns

With Agent Loop Skill

Agent loop skill triggers pattern extraction/injection:

  • On loop start: Inject relevant patterns
  • On loop completion: Extract new patterns
  • On iteration: Update pattern usage tracking

Best Practices

1. Let Patterns Stabilize

Don't prune patterns too early. Allow 5-10 uses before evaluating:

min_usage_count_for_evaluation: 5

2. Tag Patterns Consistently

Use standardized tags:

  • Language: typescript, python, go
  • Domain: auth, validation, testing
  • Type: null-safety, async, error-handling

3. Monitor Pattern Quality

Regularly review low-success patterns:

aiwg ralph-patterns list --min-success-rate 0.0 --max-success-rate 0.6

4. Share Patterns Across Teams

Export/import patterns for team-wide learning:

# Export team patterns
aiwg ralph-patterns export --output team-patterns.json

# Import on teammate's machine
aiwg ralph-patterns import --input team-patterns.json

Troubleshooting

Pattern Not Injected

Symptom: Loop doesn't receive relevant pattern

Causes:

  1. Pattern below success rate threshold
  2. Task description doesn't match semantically
  3. Injection disabled

Fix:

# Check pattern
aiwg ralph-patterns show {pattern_id}

# Lower threshold
aiwg ralph "task" --min-pattern-success-rate 0.5

# Check config
grep -A10 "cross_loop_learning:" aiwg.yml

Pattern Not Extracted

Symptom: Loop completed but no patterns added

Causes:

  1. Auto-extraction disabled
  2. Loop didn't succeed
  3. Pattern duplicates existing one

Fix:

# Extract manually
aiwg ralph-extract-patterns {loop_id}

# Enable auto-extraction
# In aiwg.yml:
ralph:
  cross_loop_learning:
    extraction:
      auto_extract_on_completion: true

Related

  • @$AIWG_ROOT/agentic/code/addons/ralph/schemas/shared-patterns.yaml - Pattern schema
  • @$AIWG_ROOT/agentic/code/addons/ralph/schemas/cross-task-memory.yaml - Cross-task memory
  • @$AIWG_ROOT/agentic/code/addons/ralph/docs/cross-loop-learning.md - Full documentation
  • .aiwg/research/paper-analysis/REF-013-aiwg-analysis.md - MetaGPT research
  • ralph skill - Iterative loop executor
  • agent-loop skill - Detects iterative loop requests

Version History

  • 2.0.0: Multi-loop awareness with loop_id tracking and cross-loop aggregation (Issue #269)
  • 1.0.0: Initial cross-task learning with semantic matching (Issues #154, #155)

References

  • @$AIWG_ROOT/agentic/code/addons/ralph/README.md — Ralph addon overview
  • @$AIWG_ROOT/agentic/code/addons/ralph/schemas/shared-patterns.yaml — Pattern schema for cross-loop learning
  • @$AIWG_ROOT/agentic/code/addons/ralph/schemas/cross-task-memory.yaml — Cross-task memory schema
  • @$AIWG_ROOT/agentic/code/addons/ralph/docs/cross-loop-learning.md — Full cross-loop learning documentation
  • @$AIWG_ROOT/agentic/code/addons/aiwg-utils/rules/subagent-scoping.md — Subagent scoping rules relevant to loop orchestration

Related skills

G

Generative Code Art

anthropics

Create algorithmic art with p5.js using randomness and interactive parameters.

OfficialComplete terms in LICENSE.txt
P

Poster & Visual Design

anthropics

Create original posters and visual art in PNG and PDF formats.

OfficialComplete terms in LICENSE.txt
C

Claude API Helper

anthropics

Build, debug, and optimize Claude API applications with caching and model migration support.

OfficialComplete terms in LICENSE.txt
M

MCP Server Builder

anthropics

Build protocol servers that connect language models to external APIs and services.

OfficialComplete terms in LICENSE.txt