blcaptain-meta-skill

BLCaptain Meta Skill is a methodology package for turning a repetitive workflow into a reusable, installable AI agent skill. The core problem it addresses: when you use an AI assistant regularly, you find yourself explaining the same process over and over. The AI does the task but forgets your preferences, ignores your standards, or produces inconsistent results each time. This project gives you a structured way to capture that process once and turn it into something the AI can reliably execute on its own.

The project is not a software library in the usual sense. It is closer to a template system and decision framework. When you invoke it, it guides you through eight stages: research, analysis, planning, development, verification, testing, audit and acceptance, and summary with iteration notes. At each stage there are defined questions to answer and deliverables to produce. The goal is a complete skill package, not just a long prompt.

A skill package in this system has a specific folder structure. The entry point is a short SKILL.md file that tells the AI when to activate, what to do first, and where to find further resources. Supporting folders hold detailed method references, templates, example runs, evaluation cases, and executable validation scripts. This thin-entry-thick-resources structure is intentional: a very long entry file tends to confuse AI agents rather than help them.

Before a workflow gets promoted to a skill, the framework runs it through a Non-Skill Gate: a check that asks whether the task is genuinely frequent, has clear output criteria, can be verified, and is worth maintaining over time. One-off questions, simple document generation, and exploratory brainstorming are explicitly out of scope.

The package is compatible with Codex, Claude Code, Claude Skills, and other agents that can read local skill folders. Installation involves cloning the repository and placing the skill folder in the appropriate directory for your chosen platform. Python validation scripts are included to verify the package structure, check route evaluation cases, and measure context budget before publishing.

Documentation is primarily in Chinese, with readme files available in ten additional languages.