vdd

Enforces evidence-based self-verification before claiming task completion or reporting review findings.

76|3|Updated Aug 7, 2026
One-click install
npx skills add https://github.com/Job-Yang/jobbyang-ai-skills --skill vdd-job-yang
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: vdd
Source: https://github.com/Job-Yang/jobbyang-ai-skills/tree/main/skills/vdd
Command: npx skills add https://github.com/Job-Yang/jobbyang-ai-skills --skill vdd-job-yang

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? AI agents tend to declare "done" or "no problem" without actually verifying their work, and when asked to review large changes they often manufacture issues or over-edit. This Skill installs a disciplined verification loop that stops premature completion claims and prevents drifting review standards. ## Core Features & Use Cases - Single-pass self-proof: Before claiming completion, expose hidden assumptions, attach graded evidence to every action, and pass four gates (time match, scope match, mechanism explanation, counter-evidence). - Anti-drift standards: Establish a three-layer ruler (system constraints, task contract, step acceptance) before starting large or multi-round work, then judge every review finding against it with only three allowed verdicts. - Use Case: After a large refactor spanning thousands of lines, use this Skill to review the change against a written task contract, reporting only deviations backed by file and line evidence instead of inventing problems to appear diligent. ## Quick Start Ask the agent to run the VDD verification loop on its last change before saying it is done.

Frequently Asked Questions about vdd

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I stop an AI agent from claiming work is done without verifying?

Apply a verification loop that requires the agent to expose hidden assumptions, attach evidence to every claim, and pass four gates covering time, scope, mechanism, and counter-evidence before stating completion. The agent must also switch to a fresh reviewer identity rather than re-walking its original reasoning.

How to review large code changes with AI without over-editing?

Write a task contract before starting that defines core goals, design decisions, and explicit no-change boundaries. During review, allow only three verdicts: matches the contract, deviates with evidence, or the contract lacks coverage, which blocks open-ended fault-finding.

Does verification-driven development require external tools or test frameworks?

No, this method runs entirely through the agent's own reasoning discipline with no external tools, scripts, or platform integration. Its effectiveness comes from the structured loop itself, making it usable in any conversation context.

When should I not run a full verification pass?

Match verification depth to the cost of being wrong. Trivial changes like copy edits need only a quick glance, while core logic, authentication, payments, or data writes justify the full multi-step verification with a switched reviewer identity.

Why does an AI agent invent problems during code review?

Agents are trained to be helpful, so a prompt framed as find issues pressures them to manufacture findings even when the design is sound. Anchoring every reported problem to a written standard prevents fabricated issues and stabilizes review scope.