Skip to content

ABOUT.md

Latest commit

 

History

History
85 lines (57 loc) · 2.81 KB

ABOUT.md

File metadata and controls

85 lines (57 loc) · 2.81 KB
Raw

About VY Prompt Master

A comprehensive prompt engineering framework for safe, deterministic AI automation.

What is VY Prompt Master?

VY Prompt Master is a prompt engineering framework designed for Vy (Vercept) – an AI-powered macOS automation agent. It transforms high-level user task descriptions into detailed, executable YAML specifications that Vy can safely and predictably carry out.

Why This Exists

Modern AI agents can perform complex computer tasks, but without proper structure, they risk:

  • Unpredictable behavior – Actions may vary between runs
  • Safety issues – Destructive actions without user consent
  • Debugging difficulties – Hard to trace what went wrong
  • Hallucination – Claiming completion without verification

VY Prompt Master solves these problems by enforcing a rigorous specification format where every action is:

  • Grounded in UI – Tied to visible elements, not abstract concepts
  • Verified – Every action has observable success criteria
  • Reversible-first – Preferring undo-able actions with safety gates
  • Deterministic – Same input produces same output

Key Features

Feature Description
Policy Router Classifies requests as allowed/disallowed/ambiguous
UI Grounding Actions reference specific, visible UI elements
Safety Gates Irreversible actions require explicit confirmation
Fallback Paths Alternatives when primary approach fails
Schema Validation JSON Schema ensures specification consistency
Failure Playbooks Predefined handling for common errors

Who Is This For?

  • Prompt Engineers building automation for Vy
  • AI Developers creating structured agent specifications
  • Automation Specialists needing deterministic UI automation
  • Organizations requiring auditable, safe AI operations

Core Philosophy

"If VY cannot verify it, VY should not execute it."

Every step follows the pattern:

locate → confirm_target → act → verify_outcome

This ensures no action proceeds without proper targeting, and no action completes without verification.

Getting Started

  1. Review the WORKFLOW.md for complete process understanding
  2. Use the ai-review-prompt.md as your AI system message
  3. Validate outputs against the schema
  4. Test with examples in the examples/ directory

Links