self-learning-claude

name

self-learning-claude

description

Self-Learning Claude (SLC) framework for building self-improving LLM systems through evolving context playbooks. Use this skill when building agents, workflows, or domain-specific systems that need to accumulate knowledge over time, when optimizing prompts/system instructions through iteration, when implementing memory systems for LLM agents, or when you need structured approaches to prevent context collapse and brevity bias in iterative LLM refinement.

SLC: Self-Learning Claude

SLC enables Claude to learn from experience by maintaining a local playbook of patterns that gets smarter over time.

How It Works

Each project directory gets a .playbook/slc-playbook.json that stores learned patterns:

~/project-a/.playbook/slc-playbook.json  # Patterns for project A
~/project-b/.playbook/slc-playbook.json  # Patterns for project B

Automatic Workflow

1. At Task Start - Load Patterns

python3 /path/to/slc_learn.py load --endpoint "METHOD /path"

2. After Success - Record What Worked

python3 /path/to/slc_learn.py success \
    --helpful "P-abc123,S-def456" \
    --lesson "What you discovered" \
    --category strategy \
    --endpoint "METHOD /path"

3. After Failure - Record What Went Wrong

python3 /path/to/slc_learn.py failure \
    --lesson "What failed and why" \
    --endpoint "METHOD /path"

Categories

Category	Use For
pitfall	Errors, failures, things to avoid
strategy	Approaches that worked well
domain	Project-specific terminology mappings
endpoint	Specific to one API/function
code	Code patterns, commands, workarounds

CLI Reference

# Load patterns for current task
slc_learn.py load [-e "endpoint"]

# Record success
slc_learn.py success [--helpful ids] [--lesson text] [--category cat] [-e endpoint]

# Record failure
slc_learn.py failure [--harmful ids] [--lesson text] [-e endpoint]

# Initialize with seed patterns
slc_learn.py seed [--force]

# Show statistics
slc_learn.py stats

Good vs Bad Lessons

Good (specific, actionable):

• "API endpoint requires array type for 'items' field, not object"
• "Authentication token expires after 1 hour - implement refresh logic"
• "Database connection must be closed explicitly in finally block"

Bad (vague, generic):

• "Be careful with types"
• "Check the schema"
• "This is tricky"

Integration

Add to your CLAUDE.md:

## Self-Learning Mode

Use SLC learning for this project. At task start run `slc_learn.py load`,
after success run `slc_learn.py success --lesson "..."`,
after failure run `slc_learn.py failure --lesson "..."`.

The script location should be referenced from your project's CLAUDE.md.