gnkbhuvan/ai-engineering-gates

This repository provides four structured skill files that guide an AI coding agent to slow down and think before it acts. The core idea is that AI language models are text-completion engines: they produce plausible output without knowing if it is correct. These four skills inject decision trees into the agent's workflow to force it to ask the right questions before writing any code or output. The four skills cover prompt engineering (clarify requirements before writing prompts), agentic AI (question whether an agent architecture is even needed before building one), FastAPI and generative AI (load models once at startup, go fully async, type everything with Pydantic), and production RAG (decide whether retrieval-augmented generation is the right approach before paying for embedding and storage). Each skill file is a gated checklist rather than a tutorial: the agent follows the decision tree, checks each gate, and only proceeds when each condition is met. The repository also includes a benchmark suite that validates the skills actually change agent behavior. Tests run without any API calls for the self-test phase, then compare real agent outputs with and without each skill loaded. The methodology uses 8 tasks across 2 test arms, scored by a mix of deterministic checks, behavioral probes, and an LLM-as-judge rubric. To use it, you clone the repo and point your AI agent at the relevant SKILL.md file. It works with Claude Code, Cursor, Windsurf, Codex, Gemini CLI, and about 20 other agents through a one-line install command or a manual clone. The agent reads the decision tree in the skill file and loads extra reference documents only when a specific branch of the tree requires them, keeping context use low. This is a developer-facing toolkit for people building or using AI coding agents who want structured, measurable guardrails rather than vague best-practice guides.