@adlc/pi

This Skill addresses the lack of structured gates and anti-model-failure mechanisms for AI coding agents during development. In practice, AI models are prone to premature satisfaction, sycophancy, context corruption, and confident hallucinations, leading to uncontrolled code quality. While traditional SDLC defends against human shortcomings, ADLC redesigns gates for each phase to specifically counter model-specific failure patterns.

Usage is straightforward: simply install the @adlc/pi package in your pi terminal and invoke native tools like adlc_prosecute and adlc_gate via natural language commands. For example, tell the AI "Run a preflight check on the current code changes," and it will automatically execute the preflight package for preliminary checks; say "Review this code with prosecute," and it will call the prosecute tool for evidence verification. All tools adhere to the .adlc/ directory convention, forming a unified chain of gate evidence.

It is particularly well-suited for backend development, full-stack development, and agent development teams requiring rigorous quality control. Teams already using Claude Code for coding but struggling with unstable AI output can leverage ADLC's tools—such as model routing, behavioral drift detection, and merge prediction—to intercept issues in real-time during coding, rather than fixing them post-facto.

We recommend consistently using this toolset when starting new feature development or performing code reviews. Note that it primarily targets Node.js environments (>=18), and all tools are zero-dependency CLIs, requiring a terminal environment capable of running npm global packages. For non-Node projects, ADLC's specifications can still be referenced, though some tools (e.g., spec-lint) may require adaptation.