AI is removing the middle class of software engineering?

TL;DR Highlight

AI made code cheap but understanding is still slow, so only engineers with judgment gain value

Who Should Read

A senior developer or tech lead who reviews AI-generated giant PRs daily and worries about codebase quality

Core Mechanics

AI removed the speed limit
- The author says projects with weak engineering culture fail much faster with AI.
- In the article's example, one PR is +24506 -3938 lines and the review queue grows from 7 to 13 PRs.
Code got cheap, understanding did not
- Producing large changes is now cheap and fast, but understanding them is still slow.
- The author says there is no shortcut yet for building a correct mental model of a change.
Bad decisions are hard to revert
- An LLM can add DB tables and columns in about 10 minutes, but once data is stored you can't just remove them.
- Migration plans, failure handling and orphaned foreign keys are still human work, he says.
Existing process wasn't built for this speed
- Tests, CI and code review were designed for a world where massive change was impossible.
- With validation as the bottleneck, he sees three options: generate less, find better validation, or accept lower quality.

Community Reactions

AI amplifies bad engineers 10x, one view goes
- Commenters said long-tenured engineers who lost interest are especially dangerous.
- Another view: AI is only as good as the abstractions and contracts you give it.
Fighting Kafka and infra changes weekly, one commenter reports
- They spend hours a week fighting people adding infrastructure layers they don't understand.
- Higher change velocity turned it into whack-a-mole.
Never using full-auto mode, one commenter says
- They don't approve anything they don't understand.
- They study unfamiliar parts first, then approve, deny or redirect.
Counterpoint: nobody explained things before AI either
- Several commenters said original authors had left and they had to dig into the code alone.
- They asked why relying on an agent for that is necessarily worse.
If AI can't fix a bug and fixes create more, the design is bad, one view goes

How to Apply

Add a self-explanation section to your PR template
- Today, add fields where the author writes data sources, why this structure, and why each new abstraction.
- If it's blank and only a Claude link is given, don't start reviewing.
Set a PR size cap
- Agree on a line or file limit and send oversized PRs back to be split.
- Tell agents to work in small pieces too.
Require design notes for new infra and schema changes
- For Kafka, serverless or new tables, get the justification plus migration and rollback plans first.
- Don't merge changes that came code-first without reasons.
Measure team throughput, not PR count
- Track review time and reverted changes alongside PR counts, and share it with management.
- Stop using metrics that show only PRs and lines changed.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the article's 'verify understanding' idea)

You are a demanding senior reviewer. From the PR description and diff below, write questions that test whether the author understands the change.

{{PR description}}

{{diff}}

Output the following.
1. A list of newly introduced abstractions, infra and schema changes, with a question about why each is needed
2. Three questions the author must answer themselves about where the data comes from and goes
3. Questions about hard-to-revert changes (migrations etc.) and the failure plan
4. A proposal to split this PR into 3 or fewer smaller PRs

Next step: post the questions as real PR comments and have the author answer without AI.

Terminology

denormaliseStoring the same data redundantly across tables to make reads easier. It looks convenient but data drifts out of sync later.

blast radiusHow far the damage from one mistake spreads, like the radius of an explosion.

vibe coderSomeone who works by throwing prompts at an agent and taking the result without really understanding the code.

orphaned foreign keyA foreign key whose target row is gone. Like mail still being sent to an address someone moved out of.