Asked claude code to verify features in a moderately large codebase. The capability is insane, so is the token usage.

I've been using AI for code reviews for over a year. In that time I've learned one uncomfortable truth:

Most developers are using AI wrong for code reviews.

They paste code and ask "is this good?" They get back a wall of generic feedback that could apply to literally any codebase. It feels useful for about 10 seconds, then you realize nothing actionable came out of it.

The problem isn't the AI. It's the prompt.

After hundreds of iterations, I've identified the patterns that separate a mediocre AI code review from one that actually finds bugs, catches security holes, and suggests fixes a senior engineer would be proud of.

Here's what I learned — and the exact prompts I now use daily.

Released May 28, the Claude Opus 4.7 upgrade beats its predecessor, GPT-5.5, and Gemini 3.1 Pro across almost all benchmarks. Mythos 1 and Sonnet 4.8 could be next.

The 7 AI coding skills I use every single day.

(All free to download):

If you spend any time in AI circles online, it's easy to come away thinking you need hundreds of skills, dozens of plugins, and an ever-growing stack of MCP servers to be productive with coding agents.

I've come to believe the opposite.

The engineers I see shipping the most consistent, high-quality work tend to use a small number of well-designed skills that map to the workflows they repeat every day. Planning, implementing, reviewing. That's most of the job.

I've spent the last couple of years building, breaking, and rebuilding my own toolkit. It's settled into just a few skills that I genuinely use every day across professional projects, and that I'd happily defend as the only ones most engineers need.

I just put together a full video walking through all seven, with live demos in Codex (though they work fine in Claude Code or any other agent).

I show how I use each one, why it earns its place, and the pattern underneath them that I think matters more than the list itself.

https://lnkd.in/e_r62kya

Every skill is free and linked in the description so you can grab them and try them yourself.

If you've got a skill you swear by that you think I'm missing, let me know in the comments.

The best part of working in public is the steady stream of better ideas coming back from people who've solved problems I haven't noticed yet.

---

♻️ Repost if you found this useful.

A practical look at Anthropic’s agentic coding tool — what it actually does, how it changed my workflow, and whether it’s worth your time.

I’m a Principal Developer and I haven’t written a line of code in a year.

That’s a strange sentence to write.

A year ago, I was still deep in C#, TypeScript, APIs, infrastructure, architecture reviews, debugging production systems, Terraform, and CI/CD pipelines.

Today?

I mostly describe systems.
I talk to AI.
I architect with AI.
I review with AI.
I direct, refine, test, challenge, and iterate with AI.

But physically typing code?

Almost never.

The last thing I manually “coded” was tweaking a bit of Terraform. Even that now feels one voice command away from disappearing entirely.

And honestly, it’s unsettling.

I genuinely feel like an accountant in 1863 who’s just been handed a MacBook Pro and a subscription to Xero.

Not because it’s impossible to comprehend.

Because within minutes you realise entire industries are about to change around it.

And then the terrifying thought arrives:

What could somebody from that era have built if they’d truly understood the tool they were holding?

That’s the uncomfortable part about the current AI wave.

Not the hype.

Not the demos.

The speed.

Because we’re rapidly moving toward a world where a non-technical person says:

“I want a CRM system that connects warehouse operations, customer service, complaints, sales, marketing, IT, security testing, and technical teams” and I want it to solve operational problems.

And increasingly, the answer is no longer:

“That will take a team of developers 18 months.”

The answer is:

“Okay.”

That’s the shift.

Not years away.

Months away if not days.

Software development itself is becoming abstracted.

The value is moving higher up the stack:

Understanding systems
Understanding businesses
Understanding people

I’m obsessed with AI because I understand what it can deliver.

The closer you are to the technology, the less theoretical it feels.

I sit there sometimes thinking:

What do you even tell your children to learn now?

What skills still compound?

What does society look like in 18 months if this pace continues?

For decades we built society around knowledge accumulation.

Go to university.
Build expertise.
Become specialised.

But what happens when intelligence itself becomes massively accessible?

What happens when execution collapses from years into days?

It’s beginning to feel like the bottleneck is no longer software development.

Delivery is rapidly becoming commoditised.

The people who win over the next few years probably won’t be the people who produce the most output manually.

They’ll be the people who can identify valuable problems and direct intelligence effectively.

That’s partly why I’m so focused on AI now.

Because it feels inevitable.

And honestly, the biggest challenge no longer feels technical.

The challenge is figuring out where to apply all of this capability before the rest of the world catches up.

Because for the first time in my career, I’m not sure where the ceiling is anymore.

And I’m not sure anybody else does either.

Code review is moving into the IDE.

As agents help engineers write more code, review and verification are becoming the bottleneck. You can open a PR faster than ever, but shipping still depends on knowing whether the change is correct, safe, and ready for production.

Today, we’re bringing two review workflows to Windsurf 2.0:

1. Devin Review in Windsurf
Run Devin Review directly from the IDE. See bugs, comments, and findings on your PR without leaving your editor. Devin Review helps with deeper PR review: smart diff organization, copy/move detection, clear explanations, and a path from findings to fixes through Devin.

2. Quick Review
Run a fast local review on your working tree before opening a PR. Quick Review is powered by SWE-check, our specialized bug detection model that’s up to 10x faster than a deep review agent while preserving accuracy.

Use Quick Review before pushing. Use Devin Review when your PR is ready for deeper review.

Both are available today in Windsurf 2.0.

I was skeptical at first.

Not about AI in general — but about whether it would actually fit into my workflow. I work mostly with legacy PHP and jQuery. The kind of codebase that was written before half the frameworks people talk about today even existed. Some Vue.js here and there in newer parts, but a lot of the core is raw PHP, procedural logic, and jQuery doing things you probably don't want to know about.

It's not glamorous. But it's real work, and it has its own kind of complexity.

When I started using Claude more seriously, I wasn't expecting much. Turns out I was wrong.

I had a draft post sitting in my local repo for a while, where I was about to scream about how AI is overestimated. Well, that post aged pretty badly. I never published it, and looking back at the notes I’m glad I didn’t. So what I’m going to write today will only be about my current workflow and how I actually use AI in my daily work — no hype, no predictions, just what I’ve found useful.

About 15 months ago I left a comfortable senior engineering role at a fintech company to go independent and build software products with AI coding assistants as my primary collaborators. I'm 38, have a mortgage, and my wife was pregnant with our second kid at the time. Not exactly the ideal moment for a career experiment.

I want to share what that experience has actually been like, because there's a lot of hype and doom out there, and not enough honest accounts from people who've spent serious time in the trenches with these tools.

Background

I've been writing software professionally for about 16 years. Mostly backend -- Java, Python, some C++ earlier on. I'm not a 10x developer. I'm a pretty average senior engineer who got tired of sprint planning meetings and wanted to build things on my own terms.

I committed to using AI coding assistants for everything. I rotate between a few different ones depending on the task -- they all have different strengths and keep leapfrogging each other every few months.

What 1,800 hours looks like

I tracked my time carefully because I'm billing myself against savings. Roughly 1,800 hours of active AI-assisted development this year. About 6-7 hours a day, six days a week.

I shipped three products: a multilingual document processing pipeline, a monitoring tool for small SaaS companies, and a real-time audio processing app still in beta.

Two of those required significant Rust and Go code. I had never written production Rust before this year. The AI assistants didn't just help me write unfamiliar languages -- they helped me understand the idioms, memory models, ecosystem tooling. Zero to shipping production Rust in about three months. That would have been 12-18 months solo.

I also went deep on vector embeddings, fine-tuning smaller language models, building custom data pipelines. A year ago I couldn't have explained cosine similarity. Now I have opinions about chunking strategies.

The part nobody talks about

Here's where I push back on the pure optimism narrative.

AI assistants are confident. Relentlessly, dangerously confident. They generate code that looks perfect, passes review, and has a subtle bug that surfaces three weeks later at 2am. I've lost entire days to AI-introduced issues I trusted too quickly.

I fell into what I call "velocity addiction." Moving so fast you skip careful review. You trust the output because it's been right fifteen times. Then time sixteen bites you hard.

One painful incident: an AI-generated database migration looked correct, passed tests, then corrupted two days of user data in staging. The logic error was subtle -- null handling that was technically valid but semantically wrong for my schema. Caught it before production, but it shook me.

These tools also make you feel more competent than you are. I wrote Rust that compiled and ran, but a friend with five years of Rust experience pointed out I was fighting the borrow checker in ways that would break at scale. AI helped me get it working but didn't teach me to think in Rust. There's a difference.

What I believe now

AI coding assistants are genuinely transformative for experienced developers. The key word is experienced. You need enough background to evaluate output, to smell when something's wrong even if it compiles.

The best mental model: you're directing a very talented but very junior team that never sleeps. They produce enormous amounts of work and know trivia about every framework. But they have no judgment. They don't understand your users or your architectural decisions. They will confidently lead you off a cliff if you let them.

The "AI will replace developers" framing is wrong, but not for comforting reasons. It's not that AI can't code -- it clearly can. It's that the hard part of software engineering was never the coding. It's figuring out what to build and why. AI is exceptional at mechanical parts and bad at strategic parts. For now.

The honest numbers

Am I more productive? Yes. Roughly 3-4x in raw output. But my error rate is higher too. I ship faster and fix more bugs. The net is positive, but it's not the clean 10x story that makes good tweets.

Am I making money? Barely. AI made building dramatically easier but didn't help with finding customers at all.

Would I do it again? Without hesitation. This year taught me more than the previous five combined.

Still figuring it out

Some days I feel like I'm living in the future. Other days I'm mass-reverting AI-generated commits at midnight questioning my life choices.

My advice if you're leaning into this: do it, but don't trust it. Build review habits before you build velocity. Keep an honest log of time spent fixing AI-introduced issues -- that number is higher than you think.

Curious to hear from others who've spent serious time with these tools. Where do you draw the line between AI-assisted and AI-dependent?