Apply Learnings Skill

Analyze Claude Code session history to extract learnings that would have been helpful if provided earlier, then persist them for future sessions.

Overview

This skill performs comprehensive analysis of session transcripts to identify:

1. Code Pattern Learnings - Framework conventions, idioms, and best practices discovered during the session
2. Architectural Preferences - Where validation/logic should live, layer responsibilities
3. Tool Usage Patterns - Failed tool calls, user corrections, successful retries
4. Missing Context - Information the user had to provide that could be documented upfront

Usage

Analyze Current Session

python3 ~/.claude/skills/apply-learnings/scripts/analyze_session.py --scope current

Analyze All Historical Sessions

python3 ~/.claude/skills/apply-learnings/scripts/analyze_session.py --scope all

Analyze Specific Project

python3 ~/.claude/skills/apply-learnings/scripts/analyze_session.py --project /path/to/project

Workflow

When the user invokes this skill:

Step 1: Ask User for Scope

Ask which scope to analyze:

• Current session - Only the current conversation
• Current project - All sessions for the current working directory
• All sessions - Complete history across all projects

Step 2: Run Analysis

Execute the analysis script:

python3 ~/.claude/skills/apply-learnings/scripts/analyze_session.py --scope <scope>

Step 3: Present Findings

Present findings organized by category:

Code Pattern Learnings

Conventions and idioms specific to the codebase/framework:

• Dependency injection patterns (e.g., "use glue logger from context, not DI")
• Error handling conventions
• Naming patterns
• Testing patterns

Architectural Preferences

Where different concerns should be handled:

• Validation location (e.g., "validate at handler/mapper, not in controllers")
• Error transformation boundaries
• Layer responsibilities

Tool Usage Improvements

From the original tool-self-improver:

• Failed tool calls and corrections
• Common error patterns
• Successful retry patterns

Missing Context

Information that would have helped earlier:

• Project-specific conventions not in CLAUDE.md
• Framework quirks
• Team preferences

Step 4: Classify Destination for Each Learning

For each learning, determine the best destination. Do NOT put everything in CLAUDE.md — route to the most specific location: