a5c-ai/continuous-learning

Continuous Learning

Overview

Continuous learning pipeline adapted from the Everything Claude Code methodology. Automatically extracts patterns from development sessions, evaluates them with confidence scoring, and converts high-quality patterns into reusable skills.

Learning Pipeline

1. Pattern Extraction

• Analyze code changes and implementation approaches
• Identify recurring patterns and conventions
• Extract architectural decisions with rationale
• Capture error resolution strategies
• Record tool usage patterns
• Assign initial confidence scores (0-100)

2. Pattern Evaluation

• Score generalizability (0-100): cross-project applicability
• Score reliability (0-100): validation frequency
• Score impact (0-100): outcome improvement
• Composite: generalizability * 0.3 + reliability * 0.4 + impact * 0.3
• Filter below confidence threshold (default: 75)
• Merge similar patterns

3. Skill Creation

• Convert high-confidence patterns to SKILL.md format
• Write clear instructions with phases
• Include when-to-use and when-not-to-use sections
• Add usage examples and agent references
• Follow kebab-case naming convention

4. Organization

• Categorize: language-specific, domain, business, meta
• Resolve naming conflicts
• Update indexes and manifests
• Create dependency graphs

5. Version and Export

• Assign semantic versions by maturity
• Create portable export bundles
• Include usage examples and test cases
• Generate import instructions

Strategic Compaction

• Analyze context token usage
• Identify low-value context for compression
• Archive completed phases to memory files
• Calculate token savings per suggestion

When to Use

• End of development sessions
• After significant code reviews
• After debugging sessions
• Periodically during long sessions

Agents Used

• continuous-learning (custom agent for this skill)
• context-engineering (compaction analysis)