How does Claude capture feedback autonomously?

Claude monitors conversation for triggers like 'thumbs up', 'mistake', or 'that worked' and automatically executes a local script to log the feedback context without user intervention.

Can I customize the categories for my specific project?

Yes, the skill supports domain-specific tags like 'timer-logic' or 'redux-state' that help categorize feedback based on the unique requirements of your application.

How does this improve Claude's performance over time?

At the start of each session, a hook queries past failures and successes, reminding Claude of previous mistakes to avoid and patterns that worked, creating a persistent learning loop.

Where is the feedback data stored?

Feedback is stored locally within your project's .claude/memory/feedback directory. It is intended to be excluded from git to keep your logs private.

RLHF Feedback Capture

Name: RLHF Feedback Capture
Author: IgorGanapolsky

byIgorGanapolsky

0•

生产力与工作流

Automatically captures and logs AI performance feedback to create a continuous, autonomous learning loop for Claude Code.

This skill empowers Claude to autonomously track its own successes and failures by monitoring conversational cues like 'that worked' or 'that's wrong.' By logging these interactions into a local, privacy-focused memory system, Claude can reference past mistakes and successes at the start of new sessions. This creates a persistent learning loop that prevents the repetition of errors, improves adherence to project-specific patterns, and ensures Claude evolves based on direct developer feedback without requiring manual command execution.

主要功能

01Support for 'Act Don't Instruct' enforcement by logging surface-level answers

020 GitHub stars

03Local-only data storage in JSONL format for privacy and security

04Autonomous feedback capture triggered by natural conversation cues

05Session-start integration to proactively query and avoid past failures

06Context-aware tagging for domain-specific categorization like logic or UI

使用场景

01Improving domain-specific accuracy in React Native projects through targeted logic tagging

02Building a project-specific knowledge base of 'what works' for future feature implementation

03Preventing repetitive coding errors by logging regressions and failed assumptions

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add igorganapolsky/random-timer rlhf-feedback

For use in Claude.ai and ChatGPT

主要功能

01Support for 'Act Don't Instruct' enforcement by logging surface-level answers

020 GitHub stars

03Local-only data storage in JSONL format for privacy and security

04Autonomous feedback capture triggered by natural conversation cues

05Session-start integration to proactively query and avoid past failures

06Context-aware tagging for domain-specific categorization like logic or UI

使用场景

01Improving domain-specific accuracy in React Native projects through targeted logic tagging

02Building a project-specific knowledge base of 'what works' for future feature implementation

03Preventing repetitive coding errors by logging regressions and failed assumptions

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add igorganapolsky/random-timer rlhf-feedback

For use in Claude.ai and ChatGPT