If you build for humans,
humans should review the prompts.

What is a Prompt & Code Review (PCR)?

A Prompt & Code Review is a structured review of the AI prompts that produced a piece of code, surfaced alongside the code as part of the normal review workflow. Where a traditional code review asks what was written, a PCR also asks how it was written: what was asked of the AI, how the developer guided it, and what decisions were made along the way.

Think of it as reviewing a commit history, but for the prompt layer. The meaningful unit of AI-mediated work is the prompt-to-code decision, not the diff.

A PCR has three parts:

1. 1.Problem framing and expected outcomes
2. 2.Key prompts and interaction modes that led to the solution
3. 3.A brief human-written rationale for the chosen approach

The goal is not an exhaustive prompt audit. Engineers self-curate two or three key prompts that shaped their solution. The review stays focused on the decisions that actually mattered.

Why human-first?

“The notion of the programmer as an easily replaceable component in the program production activity has to be abandoned. Instead the programmer must be regarded as a responsible developer and manager of the activity in which the computer is a part.”

Peter Naur, 1985

Naur wrote that in 1985. It has only become more true. AI coding tools are genuinely powerful. They write production code at a speed no human can match. But that speed creates a new problem: the decisions shaping a codebase are happening faster than teams can track, and the humans responsible for those decisions have less and less visibility into them.

Our research points to one principle: to remain autonomous is to be the author of your reasons, not merely the approver of outputs. In practice, that means three things.

• 1.You can interrupt and override the agent at any point.
• 2.The code you ship has traceable provenance, not just AI-generated reasoning.
• 3.Changes arrive in small, test-bounded increments you can actually review.

We built PCR.dev entirely with AI tools. But as long as software is built for real people, the engineers building it should be able to account for the decisions that produced it. As teams grow more AI-native and AI tools grow more capable, the review process itself becomes training. Better prompts produce better code, and seeing what works compounds over time.

The problem we found

In our study of 20 software engineers, the same pattern kept surfacing. Junior engineers using agentic AI felt productive in the short term but reported a growing loss of ownership and understanding. The feeling was consistent enough that it shows up in the exact words they used.

Senior engineers saw the same dynamic from the other side. 6 of 10 juniors in our study reported lower confidence in their ability to code without AI after a single agentic debugging session. Seniors reviewing junior prompt histories in 12-minute sessions caught over-reliance patterns, incorrect AI suggestions, and weak prompting habits they would have missed from the code alone.

The common thread was visibility. Seniors could review the output. They could not review the process. When they first saw what a prompt review could look like, they said they wanted to implement it immediately.

How we think about it

The research pointed to three practices that help teams stay in control as AI grows more capable and more autonomous.

Research-backed

We studied the problem with real engineers before building the product. The study ran in three phases with 20 engineers — 10 junior, 10 senior — across big tech, fintech, devtools, and enterprise SaaS, combining Applied Cognitive Task Analysis, an AI-assisted debugging task in Cursor, and blind reviews of junior prompt histories by senior engineers.

Feng, Dana, Bhada Yun, and April Yi Wang. “From Junior to Senior: Allocating Agency and Navigating Professional Growth in Agentic AI-Mediated Software Engineering.” doi:10.1145/3772318.3791642

Built through PCRs

PCR.dev is built using PCR.dev. Dana and Bhada use it to audit each other throughout development, which means our own prompt histories are the first test of every feature we ship.

They also use different tools. Bhada builds with Cursor. Dana builds with Claude Code. The capture agent works the same for both, and their sessions show up side by side in the same project. That is not a coincidence. It is by design. The tool you use should not determine whether you can participate in a team review workflow.

Bhada Yun

MSc Student, ETH Zürich

HCI researcher at ETH Zürich studying how people interpret and interact with AI systems. Goes mining for rocks in his free time.

bhadayun.comLinkedIn

Dana Feng

Software Engineer, Two Sigma

Software engineer and researcher based in Brooklyn. Designs for bakeries in her free time.

dana-feng.github.ioLinkedIn

Prof. Dr. April Yi Wang

Professor, ETH Zürich

Faculty at ETH Zürich, where her lab works on human-centered computing and EdTech. Pets her Shiba Inu in her free time.

aprilwang.meLinkedIn

Open source and built to grow

The prompt capture agent is open-source and will stay that way. The AI tooling landscape is moving fast: Cursor, Claude Code, Codex, Gemini CLI, and whatever comes next. We cannot build integrations for all of them alone, and we should not have to.

If you use an AI coding tool that PCR.dev does not yet support, fork the capture agent and add it. The architecture is designed to make new integrations straightforward. Contributions are welcome.