51 comments

[ 3.5 ms ] story [ 62.1 ms ] thread
@dang looks like HN ate the leading “How” in the title.
I've been trying Claude Code for a few weeks after using Gemini Cli.

There's something a little better the tool use loop, which is nice.

But Claude seems a little dumber and is aggressive about "getting things done", often ignoring common sense or explicit instructions or design information.

If I tell it to make a test pass, it will sometimes change my database structure to avoid having to debug the test. At least twice it deleted protobufs from my project and replaced it with JSON because it struggled to immediately debug a proto issue.

I’ve seen Claude code get halfway through a small sized refactor (function parameters changed shape or something like that), say something that looks like frustration at the amount of time it’s taking, revert all of the good changes, and start writing a bash script to automate the whole process.

In that case, you have put a stop to it and point out that it would already be done if it hadn’t decided to blow it all up in an effort to write a one time use codemod. Of course it agrees with that point as it agrees with everything. It’s the epitome of strong opinions loosely held.

Claude trying to cheat its way through tests has been my experience as well. Often it’ll delete or skip them and proudly claim all issues have been fixed. This behavior seems to be intrinsic to it since it happens with both Claude Code and Cursor.

Interestingly, it’s the only LLM I’ve seen behave that way. Others simply acknowledge the failure and, after a few hints, eventually get everything working.

Claude just hopes I won’t notice its tricks. It makes me wonder what else it might try to hide when misalignment has more serious consequences.

I use Claude and like it, but this post has kind of a clunky and stilted style

So I guess the blog team also uses Claude

A repeated trend is that Claude Code only gets 70-80% of the way, which is fine and something I wish was emphasized more by people pushing agents.

This bullet point is funny:

> Treat it like a slot machine

> Save your state before letting Claude work, let it run for 30 minutes, then either accept the result or start fresh rather than trying to wrestle with corrections. Starting over often has a higher success rate than trying to fix Claude's mistakes.

That's easy to say when the employee is not personally paying the massive amount of compute running Claude Code for a half-hour.

Yeah my most common aider command sequence is

    > /undo
    > /clear
    > ↑ ↑ ↑ ⏎
(comment deleted)
And this is the marketing pitch from the people selling this stuff. ¯\_ (ツ)_/¯
Funny thing their recommendation to save state as claude code has still no ability for restore checkpoints (like cline has) despite being many times requested. Who are they kidding.
Oh, come on https://www.anthropic.com/news/how-anthropic-teams-use-claud... is not "for legal"

- Custom accessibility solution for family members

- The team created prototype "phone tree" systems to help team members connect with the right lawyer at Anthropic

- Team coordination tools

- Rapid prototyping for solution validation

So, not legal

Claude Code works well for lots of things; for example yesterday I asked it to switch weather APIs backing a weather site and it came very close to one-shotting the whole thing even though the APIs were quite different.

I use it at home via the $20/m subscription and am piloting it at work via AWS Bedrock. When used with Bedrock APIs, at the end of every session it shows you the dollar amount spent which is a bit disconcerting. I hope the fine-grained metering of inference is a temporary situation otherwise I think it will have a chilling/discouraging effect on software developers, leading to less experimentation and fewer rewrites, overall lower quality.

I imagine Anthropic gets to consume it unmetered internally so I they probably completely avoid this problem.

> it shows you the dollar amount spent which is a bit disconcerting

I can assure you that I don’t at all care about the MAYBE $10 charge my monster Claude Code session billed the company. They also clearly said “don’t worry about cost, just go figure out how to work with it”

Wouldn't it be a bit unusual if that wasn't true? We build our coding agent...with our coding agent... feeling like this is table steaks.
We passed a milestone with little fanfare. AI is now improving itself.
It'd be more interesting if they shared actual examples of complete prompts, CLAUDE.md files, settings and MCP servers to achieve certain things.

The documentation is good, but is kept relatively general and I have a feeling that the quality of Claude Code's output really depends on the specific setup and prompts you use.

Wonder what the context window is for Claude Code for internal users
Not as much as the $200 month account, that's for fucking certain.
> Create self-sufficient loops > > Set up Claude to verify its own work by running builds, tests, and lints automatically.

I’ve got much better at using Claude.md and plan files, etc., but it still goes off the rails so quickly when I try to get it to follow a normal TDD test/edit/build/commit workflow. It will report success, and then the unit tests will fail or the working copy is dirty or there are build errors, etc. An LLM may be great for figuring out what code to write and what weird tool incantation to run to debug something, but I am fed up enough I want to write my own agent, because all I want is a switch-statement state machine to manage workflow.

Using slash commands that describe engineering/task workflow (steps) and /clear once it’s done solves this.
I used to say that the last 10% of software takes 90% of the effort. All the 70 percents and 80 percents in that article reminded me of that.
So this style of articles is the future? Pages of unconnected bullet points that mention the word "Claude" at least 100 times. No real information and nothing to remember.
I found if you are tightly in the loop, keep the code highly modular, and are developing new functionality alongside tests, Claude works much better.
Yeah me too, keeping the code modular being a huge part of it. In the same vein is actually finishing feature flags and removing the old code, killing off truly dead code, tests, documentation… basically eliminate as much noise as you can, so that it reads as much exemplary code as possible.
The first example was helping debug k8s issues, which was diagnosed as IP pool exhaustion, and Claude helped them fix it without needing a network expert

But, if they had an expert in networking build it in the first place, would they have not avoided the error entirely up front?

Experts make mistakes too. In fact, all humans do.
My optimization hack is that I'm using speech recognition now with Claude Code.

I can just talk to it like a person and explain the full context / history of things. Way faster than typing it all out.

I've been pretty happy with the python package hns for this [1]. You can run it from the terminal with uvx hns and it will listen until you press enter and then copy the transcription to the clipboard. It's a simple tool that does one thing well and integrates smoothly with a CLI-based workflow.

[1] - https://github.com/primaprashant/hns

Anthropic goes hard on the virtue signalling but I still don’t get how anyone manages to satisfy “Customer may not and must not attempt to (a) access the Services to build a competing product or service”

So I just avoid it and generally think the whole thing isn’t serious, because nobody seems to care enough about the safety implications of building AGI with legal terms which are logically impossible to satisfy to demonstrate appropriate attention to detail (aka, yall are noobs)

That's reason enough to leave, if them having one 9 of reliability of the service is not enough.
It’s a funny post because our team wanted to use Claude Code, but the team plan of Claude doesn’t include Claude Code (unlike the pro plan at a similar price point). Found this out after we purchased it :( We’re not going to ask every engineer to purchase it separately.

Maybe before boasting about how your internal teams use your product, add an option for external companies to pay for it!

Industry leading AI models but basic things like subscription management are unsolved…

> When Kubernetes clusters went down and weren't scheduling new pods, the team used Claude Code to diagnose the issue. They fed screenshots of dashboards into Claude Code, which guided them through Google Cloud's UI menu by menu until they found a warning indicating pod IP address exhaustion. Claude Code then provided the exact commands to create a new IP pool and add it to the cluster, bypassing the need to involve networking specialists.

This seems rather inefficient, and also surprising that Claude Code was even needed for this.

This post had to have been written in a large part by their model, right?
It certainly feels so.
I could already tell that they do by watching how often and how badly their canvas integration breaks.
Why is everyone so careless about letting Claude Desktop upload their source code and/or private data to its servers? Or maybe I'm too paranoid for caring about where data go.
The codebase I am paid to work on is a dog's breakfast, largely agglomerated over many months by offshore contractors. If Anthropic includes it in their training set, it's their own funeral.
Because my employer tells me it’s fine so long as we’re using it with AWS Bedrock. Do I believe that Amazon then wouldn’t be siphoning all this as model training data? It’s a fun conspiracy…
Wholesale use of and/or dependence on Claude Code feels like a Faustian bargain from a personal and even a business point of view.
Claude Desktop is one of the buggier applications I've ever used and one of our jokes internally is that it seems like it was very clearly vibe coded.