How do I trigger a reflection manually?

You can invoke the skill directly by asking Claude to 'reflect' or 'learn from past sessions,' which initiates the full analysis and retraining cycle.

How many evaluations are needed before reflection occurs?

By default, the skill triggers a reflection for a specific task type once it has accumulated 5 or more evaluations since the last cycle, or if the task is manually queued.

What does the ALM Reflection skill actually do?

It audits your recent interactions with Claude, looks for instances where you had to correct the AI, and updates internal 'playbooks' so Claude doesn't make those same mistakes again.

What is a Task Playbook in ALM?

A playbook is a structured Markdown file containing specific guidance, best practices, and anti-patterns that Claude refers to when performing a specific category of work.

Does it retrain the actual LLM model?

No, it retrains a lightweight TF-IDF classifier used for task categorization and updates the text-based prompts (playbooks) that guide the LLM's behavior.

ALM Memory Reflection

Name: ALM Memory Reflection
Author: sfw

bysfw

0•

Productivity & Workflow

Analyzes task outcomes and user corrections to continuously optimize Claude's internal playbooks and classification accuracy.

The ALM Memory Reflection skill is the core optimization engine of the Active Learning Memory (ALM) system. It orchestrates a structured learning cycle by reviewing past session evaluations, identifying recurring error patterns, and synthesizing those insights into specialized markdown playbooks. By retraining an internal TF-IDF classifier and extracting global user preferences, this skill enables Claude to evolve with every interaction, proactively reducing repeated mistakes and aligning its behavior with specific developer requirements and project standards.

Key Features

01Confidence Scoring: Updates reliability metrics and tracks correction rates across different task domains.

02Granular Reflection Control: Supports targeted optimization for specific task types or full system-wide audits.

03Automated Playbook Synthesis: Generates task-specific instructions based on historical correction patterns.

04Cross-Type Pattern Detection: Identifies and extracts universal user preferences into a global configuration.

050 GitHub stars

06TF-IDF Classifier Retraining: Enhances task categorization accuracy using accumulated prompt fingerprints.

Use Cases

01Reducing recurring AI mistakes by codifying identified anti-patterns from past manual corrections.

02Improving the precision of task classification after a period of heavy use or new project onboarding.

03Personalizing Claude’s coding style across diverse projects by updating the global memory playbook.

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add sfw/alm reflect

For use in Claude.ai and ChatGPT

Download Skill