Learning Loop Skill

After completing significant tasks, capture structured lessons learned. Stores knowledge entries in a YAML knowledge base for future reference. Can also query existing entries by language, tool, or category.

Arguments

• •$ARGUMENTS -- One of:
  • •capture -- Interactive capture of a new learning (default if no args)
  • •query <term> -- Search existing learnings by keyword
  • •list [category] -- List all entries, optionally filtered by category
  • •stats -- Show knowledge base statistics

Phase 1: Mode Detection

Phase 2: Capture Flow

When capturing a new learning, gather the following fields:

Required Fields

Optional Fields

Context Inference

Before prompting, infer context from the current session:

• •Check recent git log for task context
• •Check current directory for language indicators
• •Pre-fill fields where possible to reduce friction

Phase 3: Storage

Store entries in .claude/config/knowledge_base.yml using this schema:

# Knowledge Base - Auto-managed by learning-loop skill
# Do not edit manually unless correcting entries
version: 1
entries:
  - id: "20260211-001"
    title: "Ruff replaces flake8+isort+black"
    category: tool-discovery
    language: python
    description: "Ruff is a single Rust-based tool that replaces flake8, isort, and black with faster execution."
    tags: [linting, formatting, migration]
    tool: ruff
    source: refactor-python
    confidence: high
    example: "ruff check --fix . && ruff format ."
    created: "2026-02-11T14:30:00Z"
    last_referenced: "2026-02-11T14:30:00Z"

ID Generation

Format: YYYYMMDD-NNN where NNN is a zero-padded sequence number for that date.

File Creation

If knowledge_base.yml does not exist, create it with the version header and an empty entries list before adding the first entry.

Phase 4: Query Flow

When querying, search across these fields: title, description, tags, tool, language, category.

# Search examples
/learning-loop query ruff
/learning-loop query "go test"
/learning-loop list antipattern
/learning-loop list tool-discovery
/learning-loop stats

Query Output

## Knowledge Base: Query Results

**Query**: {search-term}
**Matches**: {count}

| ID | Category | Title | Language | Confidence |
|----|----------|-------|----------|------------|
| 20260211-001 | tool-discovery | Ruff replaces flake8+isort+black | python | high |
| 20260210-003 | pattern | Use ruff in pre-commit hooks | python | high |

### Details

#### 20260211-001: Ruff replaces flake8+isort+black
- **Category**: tool-discovery
- **Language**: python
- **Description**: Ruff is a single Rust-based tool that replaces flake8, isort, and black.
- **Example**: `ruff check --fix . && ruff format .`
- **Source**: refactor-python

Stats Output

## Knowledge Base Statistics

**Total entries**: {count}
**Date range**: {oldest} to {newest}

| Category | Count |
|----------|-------|
| pattern | 12 |
| antipattern | 8 |
| tool-discovery | 5 |
| config-insight | 3 |

| Language | Count |
|----------|-------|
| python | 10 |
| go | 6 |
| node | 5 |
| general | 7 |

Phase 5: Reference Tracking

When a knowledge base entry is relevant to a current task, update its last_referenced timestamp. This enables staleness tracking and prioritization.

Safety Checks

• •Never delete existing entries -- append only
• •Validate YAML syntax before writing
• •Deduplicate: warn if a similar title+category+language already exists
• •Back up knowledge_base.yml before writing (copy to knowledge_base.yml.bak)
• •Maximum 500 entries per file (suggest archiving if exceeded)