Sometimes, you'll just have a really productive session with Claude Code doing a specific thing that maybe you need to do a lot of.
One trick I have gotten some milage out of was this: have Claude Code research Slash commands, then make a slash command to turn the previous conversation into a slash command.
That was cool and great! But then, of course you inevitably will interrupt it and need to do stuff to correct it, or to make a change or "not like that!" or "use this tool" or "think harder before you try that" or "think about the big picture"
... So you do that. And then you ask it to make a command and it figures out you want a /improve-command command.
So now you have primitives to build on!
Here are my current iterations of these commands (not saying they are optimal!)
Claude Code is hard to describe. It’s almost like I changed jobs when I started using it. I’ve been all-in with Claude as a workflow tool, but this is literally steroids.
If you haven’t tried it, I can’t recommend it enough. It’s the first time it really does feel like working with a junior engineer to me.
Does anyone have any usage guides they can recommend to feel this way about using Claude code, other than the OP article? I fired it up yesterday for about an hour and tried it on a couple tickets and it felt like a total waste of time. The answers it gave were absurdly incorrect - I was being quite specific in my prompting and it seemed to be acquiring the proper context, but just doing nothing like what I was asking.
E.g. I asked it to swap all on change handlers in a component to modify a use State rather than directly fire a network request, and then add on blurs for the actual network request. It didn't add use states and just added on blurs that sent network requests to the wrong endpoint. Bizarre.
I liked Claude Code when I used it initially to document a legacy codebase. The developer who maintains the system reviewed the documentation, and said it was spot-on.
But the other day I asked it to help add boundary logging to another legacy codebase and it produced some horrible, duplicated and redundant code. I see these huge Claude instruction files people share on social media, and I have to wonder...
Not sure if they're rationing "the smarts" or performance is highly variable.
I found cursor much better than Claude Code. Running Claude code it did so many commands and internal prompting to get a small thing done and ate up tonnes of my quota. Cursor on the other hand did it super quick and straight to the point. Claude code just got stuck in grep hell
I agree with the comparison to steroids, but then I've seen people go through the health issues caused by steroids so we might mean different things by that comparison.
There are some things in there that really take this from an average tool to something great. For example, a lot of people have no idea that it recognizes different levels of reasoning and allocates a bigger number of “thinking tokens” depending on what you ask (including using “ultrathink” to max out the thinking budget).
I honestly think that people who STILL get mostly garbage outputs just aren’t using it correctly.
Not to mention the fact that people often don't use Opus 4 and stay with Sonnet to save money.
> it really does feel like working with a junior engineer to me.
I agree. It reminds me of this one junior engineer I worked with who produced awful code, and it would take longer to explain stuff to him than to just do it myself, let alone all the extra time I had to spend reviewing his awful PRs. I had hoped he would improve over time, but he took my PR comments personally and refused to keep working with me. At least Claude doesn't have an attitude.
There is a VS Code extension for Claude Code. It's hardly more than a terminal window really, but that in itself is pretty handy. If you do /ide to connect up the extension it does a few things, but not yet anything resembling the Cursor diff experience (much less the Cursor tab experience, which is the reason I still use it).
It's great to see even the most hardcore developers who are not fond of change being happy with the latest releases related to AI-assisted development.
My workflow now boils down to 2 tools really - leap.new to go from 0 to 1 because it also generates the backend code w/ infra + deployment and then I pick it up in Zed/Claude Code and continue working on it.
How are people using this without getting rate limited non stop? I pay for Claude Pro and I sometimes can’t go more than 5 prompts in an hour without it saying I need to wait 4 hours for a cooldown. I feel like I’m using it wrong or something, it’s such a frustrating experience. How do you give it any real code context without using all your tokens so quickly?
Whenever I'm rate limited (pro max plan), I stop developing.
For anything but the smallest things I use claude code...
And even then...
For the bigger things, I ask it to propose to me a solution (when adding new features).
It helps when you give proper guidance: do this, use that, avoid X, be concise, ask to refactor when needed.
All in all, it's like a slightly autistic junior dev, so you need to be really explicit, but once it knows what to do, it's incredible.
That being said, whenever you're stuck on an issue, or it keeps going in circles, I tend to rollback, ask for a proper analysis based on the requirements, and fill in the details of necessary.
For the non-standard things (f.e. detect windows on a photo and determine the measurement in centimetres), you still have to provide a lot of guidance. However, once I told it to use xyz and ABC it just goes. I've never written more then a few lines of PHP in my life, but have a full API server with an A100 running, thanks to Claude.
The accumulated hours saved are huge for me, especially front-end development, refactoring, or implementing new features to see if they make sense.
For me it's a big shift in my approach to work, and I'd be really sad if I have to go back to the pre-AI area.
Truth to be told, I was a happy user of cline & Gemini and spent hundreds of dollars on API calls per month.
But it never gave me the feeling Claude code gave me, the reliability for this thing is saving me 80% of my time.
There must at this point be lots and lots of actual walkthroughs of people coding using Claude Code, or whatever, and producing real world apps or libraries with them right? Would be neat to have a list because this is what I want to read (or watch), rather than people just continuously telling me all this is amazing but not showing me it’s amazing.
I see two trends recently:
1. low skill people using it for trivial projects.
2. devs writing out the whole app into memory/context down to the names of the files, interfaces, technologies, preparing the testing, compilation framework and then hand-holding the llm to arrive (eventually) at an acceptable solution. Example: https://www.youtube.com/watch?v=Y4_YYrIKLac
80% (99%?) of what you hear about llms are from the first group, amplified by influencers.
I'm guessing people feel the productivity boost because documenting/talking to/guiding/prompting/correcting an LLM is less mentally taxing than actually doing the work yourself even though time taken or overall effort is the same. They underestimate the amount of work they've had to put in to get something acceptable out of it.
I started using Claude code (CC) a couple weeks back and have some very positive outcomes. For clarity, I have been in the IT field since 1990, and my background is mainly infrastructure engineering (now DevOps). I don't write code professionally; I write tools as needed to accomplish tasks. That said, I understand end-to-end systems and the parts in the middle pretty well.
Here are some projects Claude has helped create:
1. Apache Airflow "DAG" (cron jobs) to automate dumping data from an on-prem PGSQL server to a cloud bucket. I have limited Python skills, but CC helped me focus on what I wanted to get done instead of worrying about code. It was an iterative process over a couple of days, but the net result is we now have a working model to easily perform on-prem to cloud data migrations. The Python code is complex with lots of edge conditions, but it is very readable and makes perfect sense.
2. Custom dashboard to correlate HAProxy server stats with run-time container (LXC) hooks. In this case, we needed to make sure some system services were running properly even if HAProxy said the container was running. To my surprise, CC immediately knew how to parse the HAProxy status output and match that with internal container processes. The net for this project is a very nice dashboard that tells us exactly if the container is up/down or some services inside the container are up/down. And, it even gives us detailed metrics to tell us if PGSQL replication is lagging too far behind the production server.
3. Billing summary for cloud provider. For this use case, we wanted to get a complete billing summary from our cloud provider - each VM, storage bucket, network connection, etc. And, for each object, we needed a full breakdown (VM with storage, network, compute pricing). It took a few days to get it done, but the result is a very, very nice tool that gives us a complete breakdown of what each resource costs. The first time I got it working 100%, we were able to easily save a few thousand $$ from our bill due to unused resources allocated long ago. And, to be clear, I knew nothing about API calls to the cloud provider to get this data much less the complexities of creating a web page to display the data.
4. Custom "DB Rebuild" web app. We run a number of DBs in in our dev/test network that need to get refreshed for testing. The DB guys don't know much about servers, containers, or specific commands to rebuild the DBs, so this tool is perfect. It provides a simple "rebuild db" button with status messages, etc. I wrote this with CC in a day or so, and the DB guys really like the workflow (easy for them). No need to Github tickets to do DB rebuilds; they can easily do it themselves.
Again, the key is focusing my energy on solving problems, not becoming a python/go/javascript expert. And, CC really helps me here. The productivity our team has achieved over the past few weeks is nothing short of amazing. We are creating tools that would require hiring expert coders to write, and giving us the ability to quickly iterate on new business ideas.
After reading and hearing rave reviews I’d love to try Claude Code in my startup. I already manage Claude Team subscription, but AFAIK Code is not included, it only exists in Pro/Max which are for individual accounts. How do people use it as a subscription for a team (ideally with central billing)?
Yeah this is super annoying and wish they'd fix it. I missed the asterisk and bought Claude Team for my team so they could use claude code but then saw it's excluded and had to go through a refund cycle and now they have to buy it individually.
HN has flipped so quickly on saying how AI produces unreliable slop, to most people using it to replace junior devs at their org – something I was heavily criticised for saying orgs should be doing a few months back.
Progress doesn't end here either, imo CC is more a mid-level engineer with a top-tier senior engineer's knowledge. I think we're getting to the point where we can begin to replace the majority of engineers (even seniors) for just a handful of seniors engineers to prompt and review AI produced code and PRs.
Not quite there yet, of course, but definitely feeling that shift starting now... There's going to be huge productivity boosts for tech companies towards the end this year if we can get there.
My own experience is that it is below sonnet and opus 4.0 on capability - but better than gemini 2.5 pro on tool calling. It's really worth trying if you don't want to spend the $100 or $200 per month on Claude Max. I love how succinct the model is.
> you can use CC with any model via
Anthropic should just open source Claude Code - they're in a position to become the VS Code of cli coding agents.
It's great for me. I have a claude.md at the root of every folder generally, outlined in piped text for minimal context addition about the rulesets for that folder, it always creates tests for what it's doing and is set to do so in a very specific folder in a very specific way otherwise it tries to create debug files instead. I also have set rules for re-use so that way it doesn't proliferate with "enhanced" class variants or structures and always tries to leverage what exists instead of bringing in new things unless absolutely necessary. The way I talk to it is very specific as well, I don't write huge prose, I don't set up huge PRDs and often I will only do planning if its something that I am myself unsure about. The only time I will do large text input is when I know that the LLM won't have context (it's newer than it's knowledge window).
I generally get great 1-shot (one input and the final output after all tasks are done) comments. I have moved past claude code though I am using the CLI itself with another model although I was using claude code and my reason for switching isn't that claude was a bad model it's just that it was expensive and I have access to larger models for cheaper. The CLI is the real power not the model itself per-se. Opus does perform a little better than others.
It's totally made it so I can do the code that I like to do while it works on other things during that time. I have about 60-70 different agent streams going at a time atm. Codebases sizes vary, the largest one right now is about 200m tokens (react, typescript, golang) in total and it does a good job. I've only had to tell it twice to do something differently.
What do you consider an “agent stream”? I can’t even imagine the cognitive overhead of managing 60-70 agents let alone the physical ability to churn through them as they complete their work and re-launch them.
One thing I'm slightly anxious about in this new LLM world is whether the prices I'm paying are sustainable. I crank the fuck out of cursor and I think we're paying like 40 bucks a month for the privilege. Is this early Uber where it was unbelievable how cheap the rides were? In 2030 am I going to have gotten dependent on ai assisted levels of productivity, be making client and customer promises based on that expectation, but suddenly find myself looking at 1k+ bills, now that all the ai companies need to actually make money?
Can someone offer me some help? I've just been messing about "vibe coding" little python apps with local llm, continue and vscode. And I got so far with it.
Then I found the free tier of claude so I fed in the "works so far" version with the changes that the local llm made, and it fixed and updated all the issues (with clear explanation) in one go. Success!
So my next level attempt was to get all the spec and prompts for a new project (a simple manic miner style 2d game using pygame). 8 used chat gpt to craft all this and it looked sensible to me with appropriate constraints for different parts of the projrct.
Which claude created. But it keeps referring to a method which it says is not present in the code and that I'm running the wrong version. (I'm definitely not). I've tried indicating it by reference to the line number and the surrounding code but it's just gas lighting me.
Any ideas how to progress from this? I'm not expecting perfection, but it seems it's just taken me to a higher level before it runs into essentially the same issue as the local llm.
All advice appreciated, I'm just dabbling with this four a bit of fun when I can (I'm pretty unwell so do things as and when I feel up to it)
For me the best part about AI is that when I'm feeling lazy, I can tell the AI to do it. Whether it gives me gold or gives me shit, it doesn't matter, because I have now started my work.
Yes, it helps a lot overcome that initial momentum. I also find it helps at the end of the day when I’m tired. I can just tell it to do the work a little bit at a time and review/fix the output. Takes less out of me.
using claude code is absolutely incredible sometimes and absolute junk other times. great for easy to mid complexity refactors. ask it to do anything decently complex and it explodes. It's taken the place of a junior dev for me now. well worth the $200/mo.
With Cursor and CC Max both throttling and changing terms recently, I worry that the trend has started to just charge and throttle more and more like a drug dealer, until we’re all stealing and living on the floor in some abandoned building.
Reading all these glowing reviews of Claude Code, I still get the feeling that either everyone’s been paid off or it’s just the die-hard fans of terminal windows and editors like Emacs and Vim. Using the terminal is right up their alley—it’s in their DNA.
Every time I read comments saying Claude Code is far better than Cursor, I fire it up, pay for a subscription, and run it on a large, complex TypeScript codebase. First, the whole process takes a hell of a lot of time. Second, the learning curve is steep: you have to work through the terminal and type commands.
And the outcome is exactly the same as with the Claude that’s built into Cursor—only slower, less clear, and the generated code is harder to review afterward. I don’t know… At this point my only impression is that all those influencers in the comments are either sponsored, or they’ve already shelled out their $200 and are now defending their choice. Or they simply haven’t used Cursor enough to figure out how to get the most out of it.
I still can’t see any real advantage to Claude Code, other than supposedly higher limits. I don’t get it. I’ve already paid for Claude Code, and I’m also paying for Cursor Pro, which is another $200, but I’m more productive with Cursor so far.
I’ve been programming for 18 years, write a ton of code every single day, and I can say Cursor gives me more. I switch between Gemini 2.5 Pro—when I need to handle tasks with a big, long context—and Claude 4.0 for routine stuff.
So no one has convinced me yet, and I haven’t seen any other benefit. Maybe later… I don’t know.
I used Cursor daily for about six months before switching to CC and it took awhile. TUI is not my first choice. But their tools and prompts make it work fare more coherently, and I've had far fewer train wrecks in my code since fully switching a few weeks ago. The IDE extension helps.
64 comments
[ 6.4 ms ] story [ 86.7 ms ] threadOne trick I have gotten some milage out of was this: have Claude Code research Slash commands, then make a slash command to turn the previous conversation into a slash command.
That was cool and great! But then, of course you inevitably will interrupt it and need to do stuff to correct it, or to make a change or "not like that!" or "use this tool" or "think harder before you try that" or "think about the big picture" ... So you do that. And then you ask it to make a command and it figures out you want a /improve-command command.
So now you have primitives to build on!
Here are my current iterations of these commands (not saying they are optimal!)
https://github.com/ctoth/slashcommands/blob/master/make-comm...
https://github.com/ctoth/slashcommands/blob/master/improve-c...
If you haven’t tried it, I can’t recommend it enough. It’s the first time it really does feel like working with a junior engineer to me.
E.g. I asked it to swap all on change handlers in a component to modify a use State rather than directly fire a network request, and then add on blurs for the actual network request. It didn't add use states and just added on blurs that sent network requests to the wrong endpoint. Bizarre.
My own experiments only show that this technology is unreliable.
But the other day I asked it to help add boundary logging to another legacy codebase and it produced some horrible, duplicated and redundant code. I see these huge Claude instruction files people share on social media, and I have to wonder...
Not sure if they're rationing "the smarts" or performance is highly variable.
If you want to try something better than claude code, try Cline.
There are some things in there that really take this from an average tool to something great. For example, a lot of people have no idea that it recognizes different levels of reasoning and allocates a bigger number of “thinking tokens” depending on what you ask (including using “ultrathink” to max out the thinking budget).
I honestly think that people who STILL get mostly garbage outputs just aren’t using it correctly.
Not to mention the fact that people often don't use Opus 4 and stay with Sonnet to save money.
I feel like working with Claude is what it must feel like for my boss to work with me. “Look, I did this awesome thing!”
“But it’s not what I asked for…”
I agree. It reminds me of this one junior engineer I worked with who produced awful code, and it would take longer to explain stuff to him than to just do it myself, let alone all the extra time I had to spend reviewing his awful PRs. I had hoped he would improve over time, but he took my PR comments personally and refused to keep working with me. At least Claude doesn't have an attitude.
My workflow now boils down to 2 tools really - leap.new to go from 0 to 1 because it also generates the backend code w/ infra + deployment and then I pick it up in Zed/Claude Code and continue working on it.
For anything but the smallest things I use claude code...
And even then...
For the bigger things, I ask it to propose to me a solution (when adding new features).
It helps when you give proper guidance: do this, use that, avoid X, be concise, ask to refactor when needed.
All in all, it's like a slightly autistic junior dev, so you need to be really explicit, but once it knows what to do, it's incredible.
That being said, whenever you're stuck on an issue, or it keeps going in circles, I tend to rollback, ask for a proper analysis based on the requirements, and fill in the details of necessary.
For the non-standard things (f.e. detect windows on a photo and determine the measurement in centimetres), you still have to provide a lot of guidance. However, once I told it to use xyz and ABC it just goes. I've never written more then a few lines of PHP in my life, but have a full API server with an A100 running, thanks to Claude.
The accumulated hours saved are huge for me, especially front-end development, refactoring, or implementing new features to see if they make sense.
For me it's a big shift in my approach to work, and I'd be really sad if I have to go back to the pre-AI area.
Truth to be told, I was a happy user of cline & Gemini and spent hundreds of dollars on API calls per month. But it never gave me the feeling Claude code gave me, the reliability for this thing is saving me 80% of my time.
80% (99%?) of what you hear about llms are from the first group, amplified by influencers.
I'm guessing people feel the productivity boost because documenting/talking to/guiding/prompting/correcting an LLM is less mentally taxing than actually doing the work yourself even though time taken or overall effort is the same. They underestimate the amount of work they've had to put in to get something acceptable out of it.
Here are some projects Claude has helped create:
1. Apache Airflow "DAG" (cron jobs) to automate dumping data from an on-prem PGSQL server to a cloud bucket. I have limited Python skills, but CC helped me focus on what I wanted to get done instead of worrying about code. It was an iterative process over a couple of days, but the net result is we now have a working model to easily perform on-prem to cloud data migrations. The Python code is complex with lots of edge conditions, but it is very readable and makes perfect sense.
2. Custom dashboard to correlate HAProxy server stats with run-time container (LXC) hooks. In this case, we needed to make sure some system services were running properly even if HAProxy said the container was running. To my surprise, CC immediately knew how to parse the HAProxy status output and match that with internal container processes. The net for this project is a very nice dashboard that tells us exactly if the container is up/down or some services inside the container are up/down. And, it even gives us detailed metrics to tell us if PGSQL replication is lagging too far behind the production server.
3. Billing summary for cloud provider. For this use case, we wanted to get a complete billing summary from our cloud provider - each VM, storage bucket, network connection, etc. And, for each object, we needed a full breakdown (VM with storage, network, compute pricing). It took a few days to get it done, but the result is a very, very nice tool that gives us a complete breakdown of what each resource costs. The first time I got it working 100%, we were able to easily save a few thousand $$ from our bill due to unused resources allocated long ago. And, to be clear, I knew nothing about API calls to the cloud provider to get this data much less the complexities of creating a web page to display the data.
4. Custom "DB Rebuild" web app. We run a number of DBs in in our dev/test network that need to get refreshed for testing. The DB guys don't know much about servers, containers, or specific commands to rebuild the DBs, so this tool is perfect. It provides a simple "rebuild db" button with status messages, etc. I wrote this with CC in a day or so, and the DB guys really like the workflow (easy for them). No need to Github tickets to do DB rebuilds; they can easily do it themselves.
Again, the key is focusing my energy on solving problems, not becoming a python/go/javascript expert. And, CC really helps me here. The productivity our team has achieved over the past few weeks is nothing short of amazing. We are creating tools that would require hiring expert coders to write, and giving us the ability to quickly iterate on new business ideas.
Besides that, I think your best bet is to find someone on youtube creating something "live" using a LLM. Something like this: https://www.youtube.com/watch?v=NW6PhVdq9R8
Progress doesn't end here either, imo CC is more a mid-level engineer with a top-tier senior engineer's knowledge. I think we're getting to the point where we can begin to replace the majority of engineers (even seniors) for just a handful of seniors engineers to prompt and review AI produced code and PRs.
Not quite there yet, of course, but definitely feeling that shift starting now... There's going to be huge productivity boosts for tech companies towards the end this year if we can get there.
Exciting times.
care to provide any proof for that? in my experience it's neither.
The recent Kimi-K2 supposedly works great.
My own experience is that it is below sonnet and opus 4.0 on capability - but better than gemini 2.5 pro on tool calling. It's really worth trying if you don't want to spend the $100 or $200 per month on Claude Max. I love how succinct the model is.
> you can use CC with any model via
Anthropic should just open source Claude Code - they're in a position to become the VS Code of cli coding agents.
Shout out to opencode:
https://github.com/sst/opencode
which supports all the models natively and attempts to do what CC does
I generally get great 1-shot (one input and the final output after all tasks are done) comments. I have moved past claude code though I am using the CLI itself with another model although I was using claude code and my reason for switching isn't that claude was a bad model it's just that it was expensive and I have access to larger models for cheaper. The CLI is the real power not the model itself per-se. Opus does perform a little better than others.
It's totally made it so I can do the code that I like to do while it works on other things during that time. I have about 60-70 different agent streams going at a time atm. Codebases sizes vary, the largest one right now is about 200m tokens (react, typescript, golang) in total and it does a good job. I've only had to tell it twice to do something differently.
Is there a bigger disconnect on how you are judged in an interview vs the job now?
How are the AI only developers handling this?
AI can't think or reason, LLMs are still mostly useless for gaming interviews.
Then I found the free tier of claude so I fed in the "works so far" version with the changes that the local llm made, and it fixed and updated all the issues (with clear explanation) in one go. Success!
So my next level attempt was to get all the spec and prompts for a new project (a simple manic miner style 2d game using pygame). 8 used chat gpt to craft all this and it looked sensible to me with appropriate constraints for different parts of the projrct.
Which claude created. But it keeps referring to a method which it says is not present in the code and that I'm running the wrong version. (I'm definitely not). I've tried indicating it by reference to the line number and the surrounding code but it's just gas lighting me.
Any ideas how to progress from this? I'm not expecting perfection, but it seems it's just taken me to a higher level before it runs into essentially the same issue as the local llm.
All advice appreciated, I'm just dabbling with this four a bit of fun when I can (I'm pretty unwell so do things as and when I feel up to it)
Thanks in advance.
Every time I read comments saying Claude Code is far better than Cursor, I fire it up, pay for a subscription, and run it on a large, complex TypeScript codebase. First, the whole process takes a hell of a lot of time. Second, the learning curve is steep: you have to work through the terminal and type commands.
And the outcome is exactly the same as with the Claude that’s built into Cursor—only slower, less clear, and the generated code is harder to review afterward. I don’t know… At this point my only impression is that all those influencers in the comments are either sponsored, or they’ve already shelled out their $200 and are now defending their choice. Or they simply haven’t used Cursor enough to figure out how to get the most out of it.
I still can’t see any real advantage to Claude Code, other than supposedly higher limits. I don’t get it. I’ve already paid for Claude Code, and I’m also paying for Cursor Pro, which is another $200, but I’m more productive with Cursor so far.
I’ve been programming for 18 years, write a ton of code every single day, and I can say Cursor gives me more. I switch between Gemini 2.5 Pro—when I need to handle tasks with a big, long context—and Claude 4.0 for routine stuff.
So no one has convinced me yet, and I haven’t seen any other benefit. Maybe later… I don’t know.