Blog

AI Product Engineering Blog: PRDs, MCP & Code Intelligence

Specification-first AI product engineering starts from a document the model is not allowed to invent around. A PRD here is not a pitch deck and not a chat prompt. It names the user, the job to be done, the states the system can enter, the interfaces it exposes, and the constraints a coding agent has to keep. When that spec is explicit, Cursor and Claude have something to compile. When it is only a vibe, the tools guess, and the repository fills with behavior nobody reviewed.

These essays are the prd.it field notes on that failure mode. Three subjects run through the archive. Structured PRDs an LLM can follow without making product decisions on its own. The Model Context Protocol, which lets an editor read a live workspace instead of a pasted prompt. And code intelligence: visual maps, repository structure, and review, so a generated codebase stays legible to the people who own it.

The same loop is what we ship. Specifys AI writes the structured PRD before implementation starts. Rift Code connects to GitHub and turns a large repository into a map a new engineer can onboard from. GreenPRD keeps the spec on a visual canvas and serves it to the IDE over MCP, so an edit on the graph updates the document and the scaffold together. AdPulse runs the loop on creative work: product context in, on-brand ads out, and performance folded back into the next brief. See the product overview for how the four fit together.

Specifys AI

Specification-First PRD Generation

The posts in this group treat the PRD as the input a coding agent compiles, not as a document written after the code exists. They start from a failure we see in AI-assisted teams: output shows up in minutes, and the product decisions inside that output were never made by anyone. A generator will invent edge cases, permissions, and state transitions when the spec left them blank, and those inventions become the system. The essays walk through how that sprawl turns into technical debt, what a prompt architecture looks like when the requirements are structured enough for Cursor or Claude to follow, and why a specification-first workflow is the practical antidote. Read them when the model is merging faster than the team can say what should have been built.

Rift Code

Repository Intelligence & Visualization

A repository full of generated code fails in a different place: comprehension. Line diffs still render, but they stop answering the questions a reviewer actually has. Which boundary moved, which service now depends on another, and what a new engineer should look at first. These essays treat code intelligence as a product problem. One lays out a way to onboard onto a large, messy GitHub repo in minutes rather than days. Another argues that text-based review hides architectural impact. A third looks at visual architecture maps as a way to lower the cognitive load of a codebase nobody holds in their head. The spec says what should exist. The map says what does exist. The gap between them is where teams lose the product.

GreenPRD

Edit Visually, Not in Chat

Prompt engineering assumed the chat box was the durable interface, and that a person would restate the product on every turn. That assumption does not survive a long build. The Model Context Protocol is the architectural answer already shipping in editors: a standard way for a model to call tools and read resources that live outside the transcript, including the spec. The essays here explain what MCP is, how it changes day-to-day IDE workflows, and why visual workflow composition is replacing hand-written prompts as the way teams describe autonomous work. The person should edit a structure the model can read, instead of rewriting instructions every time the context window forgets. GreenPRD is that structure: a canvas, a document, and an MCP server kept in one state.