In words
What it is, why it matters, and what it is like.
Why am I learning this?
Review gates are the checkpoint that turns a specification into a controlled build. Without them, an AI coding agent can drift from your intent and you only find out when the code is wrong. Mastering review gates unlocks the rest of Spec-Driven Development: acceptance criteria you can test, traceability that catches drift early, and task cards you can delegate with confidence. They are the skill that separates 'vibe coding' from engineering.
The idea, in plain terms
Imagine you are building a house. You have a detailed blueprint, but the builders (AI agents) are fast and sometimes creative. If you never check their work against the blueprint, they might pour the foundation in the wrong place, use cheaper materials, or add a room you never asked for. A review gate is a formal stop at each stage—foundation, framing, wiring—where you, the accountable reviewer, compare the actual work to the criteria on your checklist. If it fails, work stops and gets corrected before you waste money building on top of a mistake. In software, the blueprint is the specification, and the review gate is where you—not the AI—decide if the code genuinely meets each requirement before the agent moves to the next task. This is the only point where human judgement actually enters the process; the spec does not enforce itself.
An analogy
Think of a restaurant kitchen. The head chef (you) writes a detailed recipe (the specification). The junior cooks (AI agents) prepare the dishes. But the head chef does not just hand over the recipe and go home. At each course, before it is served, the chef tastes it and compares it to the recipe: Is the salt level right? Is the sauce the right consistency? Is the plating correct? If not, the dish is sent back. This taste test is the review gate. It works because the chef has defined what 'correct' means ahead of time—salt to taste, sauce thick enough to coat a spoon—not just 'make it nice.' The review gate breaks down if the chef has no criteria: then it is just taste, and every cook will get a different opinion. It also fails if the chef delegates the tasting to a junior cook without training, because they will apply their own standards. In AI, the recipe is your specification, the cooks are the agents, and the review gate is your formal approval step. The gate works only when the criteria are explicit and when you, the accountable reviewer, do the checking.
Definition
A review gate is an explicit checkpoint in a workflow where work must be approved by a designated reviewer against predefined criteria before the process may continue.
Where this sits
You already know about engineered intent documents and product requirements from your library. Review gates build directly on those: the requirements define what 'good' looks like, and the gate enforces it. They also connect to acceptance criteria and traceability—your gate should check that each acceptance criterion is met, and that each links back to a requirement. Coming up, review gates are the foundation for test-driven development: you will learn to write tests that encode the criteria, and the gate becomes a test result.
