Hacker News Reader

Comment by jurgenburgen | Hacker News Reader

> It seems that everyone loves agentic Claude code things these days but I don’t understand how you can review what it did and remain as much in the flow as you do with cursor.

The way I review it is to let it finish a PR sized slice and then I review the whole diff in a separate terminal as if it’s a PR from someone else. Then I very impolitely list the shit that needs fixing, let the LLM grind out a new batch and repeat the process. When the code is good enough I throw it into a commit and push a PR out for other humans to review.

Clicking around in a UI would just waste time and if the change is too large to review in a terminal window the change is too large for another human to review.

View on HN

There is a step missing. Impolitely list what's wrong and try to incorporate any persistent themes into a markdown file so that the agent stops making the same mistakes. Put any rules into linters and static analysis and sanitizers and ...

I sometimes see people reviewing LLM code leaving the same review comments multiple times, or worse, after some time they just give up reviewing.

Yep, if you’re not building guardrails, using an LLM is effectively just gambling until you get a non-broken output.

I 100% agree on the automated guardrails.

The part where you add instructions into a document is a bit of hit and miss. If they’re general enough it works but at my employer they have gone all out and have LLMs writing instructions that are committed to repos. There’s also a bunch of tools enforcing security instructions that usually just make the model confused (it starts arguing against the instructions randomly in the middle of performing a task). The agents are “stuttering” on most tasks because of the bloated instructions.

Yeah, workflows like ruff then ruff --fix are ideal, or just outright banning certain libraries, or having policies like dependency cool downs.

It's funny because we're once again speed running what the best practices already are!

>The way I review it is to let it finish a PR sized slice and then I review the whole diff in a separate terminal as if it’s a PR from someone else.

Insanely inefficient. It's 10x more productive to watch the thinking traces and edits in real time and steer the model appropriately.

If your workflow is typical no wonder my team members who use claude code are so much less productive.

Do you never have agents working on multiple tasks in parallel? You'll become much more productive when you figure out how to stop micromanaging.

More productive in what way? Letting agents burn tokens to produce garbage output is not productive. And letting teammates read code that wasn't reviewed by a human submitting it also isn't productive.

If I work with one agent / few subagents on one feature I can steer it as soon as I notice it drifting into the direction of waste. This way I only review the total diff 1.5-2 times. And I also don't waste my own mental energy on context switching between tasks agents are producing diffs for in parallel.

> And letting teammates read code that wasn't reviewed by a human submitting it also isn't productive.

If that’s what you got from my comment then you need to review it again.

I don't think this is something anyone who has ever read a "thinking trace" would unironically say. Not that you can even see them in Claude.

When thinking first started and you would still see the whole "thinking process", I thought it was a ploy to 10x token use because it was just the most inane bullshit. "But wait, the user is asking me to" in loops.

You can see thinking summary transcripts in Claude Code and Desktop, and they are actually useful because they don't have those sort of thinking loops from the raw tokens.

Exactly, looking at a few of these traces is enough to realise that it's a waste of time to read them. A friend once described it as like reading a fever dream, which seems fitting. They're a necessary crutch for how the tools work, but probably should be considered an internal state representation that only sometimes accidentally seems to make sense.

Wasn't there a study recently that even found a model's performance was sometimes better when the "reasoning" was nonsense? As in, no clear correllation between what the reasoning says in a human's interpretation, and how the model actually did with the task.

I sometimes read the traces out of boredom or morbid curiosity. My favourite bit with Claude is how, even if you give a very comprehensive prompt in complete sentences, almost every trace will contain a variation of "the user asks me to X, but their thought cuts off mid-sentence."

Fever dream.

Considering the amount of rejects I do every single day I'm not sure I agree. Now that I cannot reject I am unable to do any agentic coding anymore.

You’re reviewing at the wrong layer if you’re approving / rejecting each individual tool call that modifies source code.

Look at the finished piece of work and then treat the agent as if it’s that colleague you really hate and nitpick the shit out of the change it wrote. Usually the agent writes some over-engineered garbage which you then “help it” whittle down to something you would be okay approving if another human wrote it.

Reply on news.ycombinator.com