Why reviewing AI code is a bottleneck (and how to fix it)

We’ve all been there. You give Cursor, Claude Code, or an autonomous agent a prompt. It thinks for 20 seconds and spits out a 500-line diff across 12 files.

You scroll through it. It looks like code. The tests pass. You hit "Approve" and move on.

But a recent study by JetBrains and UC Irvine found something terrifying: developers using AI assistants actually switch contexts more frequently, and 74% of developers didn't notice this increase. We are writing code faster than we can actually review it.

The cognitive load of verifying AI output is quietly accumulating, and it's leading to a new kind of technical debt: The "Blind Merge."

We don't need AI to review AI. We need a fast, deterministic way to know where to look first.

I wanted a tool that would read my git diff and just tell me: "Hey, you touched the auth folder, and you deleted two test lines. Look here first."

So I built risk-annotate.

It’s a completely local, zero-dependency CLI. No cloud, no telemetry, no API keys. It runs on any diff and annotates the riskiest hunks.

Here is what it looks like running on a recent commit:

# Risk annotations

`HEAD~1..HEAD` · 7 files · 3 HIGH · 3 MED · 1 LOW

## `src/a.test.ts`
- L1-1: [HIGH] Removes 2 test lines; confirm coverage was not dropped to make the suite pass (path heuristic)

## `src/auth/login.ts`
- L5-5: [HIGH] Touches auth code; check session, token and permission handling (path heuristic)

## `.github/workflows/ci.yml`
- L1-1: [MED] CI workflow changed; check permissions, secret usage and triggers (path heuristic)

## `package.json`
- L5-6: [MED] Adds new dependency "zod"; review provenance, license and maintenance (dependency manifest)