Agentic Feedback Studio Skill

This skill provides a production-grade Agentic Control Plane for AI workflows. It allows the agent to learn from user vibes in real-time and enforce verifiable guardrails.

Capabilities

• Vibe-to-Verification (V2V): Records up/down signals and converts them into repository-level architectural constraints (The Veto Layer).
• Agentic Guardrails: Automatically generates and enforces CLAUDE.md / AGENTS.md rules derived from recurring failure modes.
• Context Engineering: Packages high-density proprietary knowledge into "Context Packs" for improved agent reliability.
• RLHF Dataset Engineering: Exports preference pairs (Chosen vs. Rejected) for model fine-tuning.

Activation

The model should activate this skill whenever:

1. The user provides explicit feedback (e.g., "thumbs down", "that's wrong", "good job").
2. The user identifies a repeated mistake.
3. The user asks for a summary of agent performance or "what have you learned?"
4. The agent needs to verify a high-risk action against existing prevention rules.

Commands

• capture: Capture new signal.
• summary: Get performance analytics.
• rules: Sync prevention rules to the repo.
• export-dpo: Generate training data.

Environment Requirements

• Requires access to the local filesystem to read/write feedback logs in .rlhf/ or ~/.rlhf/.
• Requires MCP (Model Context Protocol) support for tool execution.