Varietyz/Disciplined-AI-Software-Development

Disciplined AI Collaboration: a method for building software with AI. One loop at every size, rules held by checks rather than attention, state derived rather than written, evidence in place of claims. Includes the PAG grammar, an architecture and a queryable ontology. Every page is also served as Markdown and JSON at banes-lab.com.

What it solves

This project provides a methodology for building software with AI to prevent common failure modes like code bloat, architectural drift, context dilution, and behavioral decay. It replaces human-remembered rules and chat-based instructions with machine-enforceable constraints and evidence-based verification.

How it works

The system operates on a continuous loop of orientation, intent, derivation, projection, action, and verification. Key mechanisms include:

  • Check-First Development: Writing verification checks before the actual code to ensure safety and consistency.
  • Tree-Based Context: Moving rules and architectural requirements from chat history into files within the project tree, which the AI reads at startup.
  • PAG (Pattern Abstract Grammar): A structured instruction format used for agents, planning templates, and validation gates to ensure deterministic execution and reduce interpretation errors.
  • The Gate: A validation process that runs once per state of the tree, checking types, dead code, linting, and custom architectural anti-patterns.

Who it’s for

Developers who use AI agents or LLMs to write the majority of their code and want to maintain long-term project stability, scalability, and architectural integrity without spending excessive time debugging AI-generated output.

Highlights

  • Adversarial Review: A stance where claims are unverified until proven by the current state of the tree.
  • Governed Autonomy: A framework where the human governs boundaries and the AI executes within them.
  • Deterministic Instructions: Use of PAG to create executable contracts for agents rather than relying on personas.
  • Evidence-Based Verification: A requirement that verification output be read in full and derived from the actual state of the tree.

Related

  • Project
    minikin/cargo-crap

    A tool for Rust projects that computes the CRAP metric to identify complex, poorly tested functions where bugs are most likely to occur.

  • Project
    huggingface/gpu-fryer

    A GPU stress-testing tool that detects thermal throttling and performance degradation across NVIDIA GPUs to ensure peak performance for ML workloads.

  • Project
    deel-ai/puncc

    A Python library for predictive uncertainty calibration and conformalization, providing rigorous uncertainty estimations for any machine learning model across various tasks.

  • Project
    feitangyuan/motion-web

    A motion-first creative web engineering skill for AI agents and developers that uses real-world physics and automated headless verification to build high-touch interactive websites.