Cursor shipped a /thermo-nuclear-code-review for the TOUGHEST AI code review possible. But is it any good? Let's dive in:

Here's the skill:

here's the full plugin btw!

How timely, I just started working on a similar skill myself (`overhaul`) for a product I'm about to launch:

thermonuclear works great on the demo codebase. the real test is running it on a 5 year old monorepo with existing debt. at that point every PR gets 50 flags and the team tunes it down until its useless. the skill that matters is triage, not strictness

This is the right kind of pressure test for coding agents: not whether they can produce a diff, but whether the review loop finds architectural risk, missing tests, and subtle regressions before humans waste time reading polished wrong code.

I think that some of these should stay in a linter that gates the code from being committed if violated and force the agent to retry. LoC, cyclomatic complexity, etc. Go deterministic on whatever you can, advice on the rest

The problem is not whether the tool is strict enough. The problem is what happens when your team sees 50 flags on a PR and nobody can agree which ones matter. I watched this play out at three different companies. The team adopts "thermonuclear" mode because it sounds disciplined. First week, the signal is clear: reject bad code, move fast. By week three, reviewers are skimming flags looking for "real" issues hidden in the noise. By month two, the team tunes down the severity to get PRs merged, and the tool becomes decoration. The trap: strictness feels like discipline. It is not. Discipline is agreeing on which rules you actually enforce, then making the tool match your standards, not the reverse. If your codebase has 15 years of debt and 200 different naming conventions, flipping "max strictness" is not an upgrade. It is betrayal by a thousand flags. The skill that matters is not in the tool. It is in the team knowing their actual quality standards first, then picking a tool that enforces exactly those standards with no more, no less. The thermonuclear version is a test, not a production mode.

Matt i'm still waiting for your /review skill my man

The thing you are missing is the recommendation system of which skill/agent/mcp to use in then. That is why i have built that layer:

@mattpocockuk honestly sounds intense. curious how it handles edge cases or if it's just smoke and mirrors. have you seen it in action yet?

@ericzakariasson Some good stuff here ^

The naming alone tells you something. Most tools call their hardest mode "expert" or "strict." Cursor went full scorched earth. I'd watch whether it actually catches architectural debt or just flags syntax patterns, because those are wildly different things.

@JhengShao

curious if "thermo nuclear" just means more volume, or if it actually gets better at triaging what matters. the best human reviewers don't find more issues, they find the right ones

It feels like this skill should be split into a few separate pieces: the first part should be baked into the coding guidelines in the coding skill, and only the second part should be used as the reviewer skill. This would avoid unnecessary duplication and force the agent to write better code from the ground up.

The useful version of "thermo-nuclear review" is not maximum criticism. It is calibrated friction: find the scary bugs, cite the evidence, and avoid turning taste into churn.

Try mine: /code-review

curious whether the review catches logic errors or mostly style/pattern violations. AI code review tends to be good at the second and miss the first

I was hopping you would of released your improved version of the skill at the end. Great review

The code reviews I actually keep are the ones that flag subtle logic drift, not lint noise. Curious if this catches the stuff that bit me on my last big refactor.

the two-axis split in your skill — standards vs does-it-match-the-spec — is the right cut, we landed on the same one. the spec axis is where the real bugs hide: clean code that passes lint and quietly solves the wrong ticket. standards automate easily; intent doesn't.

hot take: "thermo-nuclear" code reviews usually mean nobody trusted the original dev

this helped me change to a better approach on a feature I was working on, so I would say yes it’s good

One interesting thing about the thermo-nuclear-code-review skill is that it doesn't have any mention of tests or test coverage, nor does it always clearly defined what is good vs. bad.

In such cases where we want to review a codebase, is it better to have a team of sub-agents review the code, or a single agent is enough for it?

it's not quite nuclear enough am finding

Yeap I love this one. Added it as a reviewer with this exact skill for every PR/pr commit with deepseek 4 pro and it's doing spectacular job.

@grok 有youtube链接吗？

What about code-review Claude code launched couple days ago?

Strictness is the easy part. Triage is the product. A five-year monorepo will always have more issues than a team can fix, so the review has to separate mechanical checks from the few findings worth interrupting someone for.

The useful version of “brutal AI review” is not “find everything wrong” but “show me the 3 things most likely to break in production, then prove it with a repro or test” Otherwise it just becomes anxiety as a feature

Hey Matt, what do you think about codex `/review` command? The command spins up a separate agent that uses these instructions:

While skills are interesting, they are only the entry point for the agent to the context system. The real meat is in the tools which the agent has access to.

I doubt this makes the new cohort material lol