I strongly disagree. It is extremely annoying when I cannot make different agents read a different prompt file when I needed to tell them different things. Cursor currently autoloads AGENTS.md without a way to disable it and it sucks.
I use symbolic links, and Claude Code often gets confused, requiring several iterations to understand that the CLAUDE.md file is actually a symbolic link to AGENTS.md, and that these are not two different, duplicate files
The recommended approach has the advantage of separating information specific to Claude Code, but I think that in the long run, Anthropic will have to adopt the AGENTS.md format
Also, when using separate files, memories will be written to CLAUDE.md, and periodic triaging will be required: deciding what to leave there and what to move to AGENTS.md
I'm still not 100% sure I understand what a symlink in a git repository actually does, especially across different operating systems. Maybe it's fine?
Anthropic say "put @AGENTS.md in your CLAUDE.md" file and my own experiments confirmed that this dumps the content into the system prompt in the same way as if you had copied it to CLAUDE.md manually, so I'm happy with that solution - at least until Anthropic give in and support AGENTS.md directly.
It just creates the same symlink on any other checkout. (On Linux/macOS at least, Windows I believe requires local settings changes.)
Only sane (guaranteed portable) option is for it to be a relative symlink to another file within the same repo, of course. i.e. CLAUDE.md would be -> 'AGENTS.md', not '/home/simonw/projects/pelicans-on-bicycles/AGENTS.md' or whatever.
On windows, it depends on the local git configuration. It’s not something I’ve been happy with, especially since symlinks also behave differently again when you’re running a docker container to get your windows usable for development.
In my experience, neither Claude nor any other agent actually reads AGENTS.md (or CLAUDE.md or anything else) without being told to explicitly every session.
I've sniffed Claude Code's HTTP traffic and confirmed that the CLAUDE.md file content (and AGENTS.md if it is @-referenced) is automatically included in the system prompt without it having to perform any additional file read operations.
At this point these tools are all pretty good. I also feel like folks often also over index on the output style or UI. Like to me the “you’re absolutely right!” sycophancy isn’t a notable bug; it’s a signal that you’re too in-the-loop. Generally my goal is to “shoot and forget”—to delegate, set the context, and let it work. Judging the tool by the final PR and not how it gets there.
I use LLM tools daily and my experience is that this just does not work on an application of any meaningful size or complexity.
You're going to immediately be overwhelmed with code, comments, markdown files, and all manner of extra text that may or may not solve your problem. The chore then becomes PR reviews of tons of iffy code, and if you merge too much of it, after too long your codebase will be a bloated mess that even the craftiest prompts won't untangle.
It's best to use to brainstorm and implement very targeted prompts, set and forget is crazy.
> Instead of a bloated API, an MCP should be a simple, secure gateway that provides a few powerful, high-level tools [...] In this model, MCP’s job isn’t to abstract reality for the agent; its job is to manage the auth, networking, and security boundaries and then get out of the way.
Thanks! I def don't think I would have guessed this use case when MCP first came out, but more and more it seems Claude just yearns for scripting on data rather than a bunch of "tools". My/MCPs job has become just getting it that data.
Have you tried using light CLIs rather than MCP? I’ve found that CLIs are just easier for Claude, especially if you write them with Claude and during planning instruct it to think about adding guidance to users who get confused.
Our auth, log diving, infra state, etc, is all usable via cli, and it feels pretty good when pointing Claude at it.
Yeah if that's possible or you are willing to build it, that's the right solution. Today pretty much all of my integrations are pure CLIs like that rather than MCPs.
You can do anything you want via a CLI but MCP still exists as a standard that folks and platforms might want to adopt as a common interface.
Agreed. My only MCP is a code interpreter. I also recently started experimenting with making an MCP “proxy” which acts a better harness that lets the agent call MCP from within a code interpreter [1]
But in general I still don’t really use MCP. Agents are just so good at solving problems themselves. I wish MCP would mostly focus at the auth part instead of the tool part. Getting an agent access to an API with credentials usually gives them enough power to solve problems on their own.
This is how MCP works if you use it for as essential an internal tool API gateway (stateless http) instead of a client facing service that end users are connecting directly to. It's basically just OpenAPI but slightly more tuned for LLM inference.
I have the same instinctive response to reading AI generated stuff, but I'm coming to a more moderate position where I'm trying to judge the content on the content itself. For example, in a post like this, it doesn't bother me at all because it's still an extremely useful reference, and the author clearly read through, organized, and edited the output. This is a good example of usage of AI in my opinion.
The people who just copy paste output from ai and ship it as a blog post however, deserve significant condemnation for that.
Writing is a tool for thought I feel, and when you're outsourcing your thought, it detracts from whatever you intended to say. I guess if it was that heavily-edited such a telltale sign wouldn't have remained.
I use AI for code, but I never use it for any writing that is for human eyes.
Nothing there implies I was forced. I read the whole thing. It was not disclosed that it was written (in part?) by AI. And That sentence I quoted was pretty far along.
> It was not disclosed that it was written (in part?) by AI
Who gives a shit?
If you can’t stand AI writing and you made it pretty far along before getting upset, who are you mad at, the author or yourself? Would you be happier if you found out this was written without AI and that you were just bad at detecting AI writing?
> /clear + /catchup (Simple Restart): My default reboot. I /clear the state, then run a custom /catchup command to make Claude read all changed files in my git branch.
I've found myself doing similar workarounds. I'm guessing anthropic will just make the /compact command do this instead soon enough.
I've found the latency of /compact makes it unusable. Perhaps this is just the result of my waiting until I have 0% context remaining.
Fun fact, a large chunk of context is reserved for compaction. When you are shown that you have "0% context remaining," it's actually like 30% remaining that's reserved for compaction.
And yet, for some reason I feel like 50% of the time, compaction fails because it runs out of context or hits (non-rate) API limits.
Weirdly, I’ve found that when that happens I can close Claude and then run `claude --continue` and now it has room to compact. Makes no sense.
But I have no idea what state it will be in after compact, so it’s better to ask it to write a complete and thorough report including what source files to read. Lot more work but better than going off the rails.
Kinda sad if 3000 words is now considered "too long to read through rather use as reference" but some interesting points, I'd be keen to see an even longer version with actual examples instead of placeholder ones.
> If you’re not already using a CLI-based agent like Claude Code or Codex CLI, you probably should be.
Are the CLI-based agents better (much better?) than the Cursor app? Why?
I like how easy it is to get Cursor to focus a particular piece of code. I select the text and Cmd-L, saying "fix this part, it's broken like this ____."
I haven't really tried a CLI agent; sending snippets of code by CLI sounds really annoying. "Fix login.ts lines 148-160, it's broken like this ___"
Better? Hard to say. Different? Yes. Worth evaluating? Absolutely. Using it for 30 minutes will answer your question better than any reply here. I think you'll answer your own question quickly.
I've been coding seriously for about 15 years. No single tool has changed how I code more than claude code and I'm including non-"AI" tooling/services. This sounds like I'm shilling but I am not affiliated. It's played a large part in injecting my passion back into building stuff.
Yeah I started with Cursor, went hybrid, and then in the last month or so I've totally swapped over.
Part of it is the snappy more minimal UX but also just pure efficacy seems consistently better. Claude does its best work in CC. I'm sure the same is true of Codex.
All I can say is when I switched from Cursor to Claude it took me less than 24 hours to realise I wouldn’t go back. The extra UI Cursor slaps on to VS Code is just bloat, which I found quite buggy (might be better now though), and the output was nowhere near as good. Maybe things have improved since I switched but Claude CLI with VS Code is giving me no reasons to want to try anything else. Cursor seemed like a promising and impressive toy, Claude CLI is just a great product that’s delivering value for me every day.
That particular part is the same, roughly. The bigger issue is just that CC's a better agent than Cursor, last I checked.
There's even an official Anthropic VS Code extension to run CC in VS Code. The biggest advantage is being able to use VS Code's diff views, which I like more than in the terminal. But the VS Code CC extension doesn't support all the latest features of the terminal CC, so I'm usually still in the terminal.
Direct use of Codex + GPT5 or Claude Code CLI gives a better result, compared to using the same models in Cursor. I've compared both. Cursor applies some of their augmentation, which reduces the output size, probably to save on tokens.
Really, the interface isn't a meaningful part of it. I also like cmd-L, but claude just does better at writing code.
...also, it's nice that Anthropic is just focusing on making cool stuff (like skills), while the folk from cursor are... I dunno. Whatever it is they're doing with cursor 2.0 :shrug:
There's something in the prompting, tooling, heuristics inside the Claude Code CLI itself that makes it more than just the model it's talking to and that becomes clear if you point your ANTROPHIC_URL at another model. The results are often almost equivalent.
Whereas I tried Kilo Code and CoPilot and JetBrain's agent and others direct against Sonnet 4 and the output was ... not good ... in comparison.
I have my criticisms of Claude but still find it very impressive.
Man, I just want to point out how different programming has become in 2 years.
We are literally having to play mind games with our tooling now to convince and coax it into doing certain things. I swear we are not far off from having to call in sick to work because your LLM "is going through a lot right now and just needs a mental health break".
As fascinating as these tools can be - are we (the industry) once again finding something other than our “customer” to focus our brains on (see Paul Graham’s “Top idea in your mind” essay)?
It seems so ... LLM-based coding tools are mostly about speed and cost of development - corporate accounting metrics, but what customers care about is mostly product features (& lack of bugs).
There is no customer advantage to developing cheap and fast if the delivered product isn't well conceived from a current and future customer-needs perspective, and a quickly shipped product full of bugs isn't going to help anyone.
I think the same goes for AI in general - CEOs are salivating over adopting "AI" (which people like Altman and Amodei are telling them will be human level tomorrow, or yesterday in the case of Amodei), and using it to reduce employee head count, but the technology is nowhere near the human level needed to actually benefit customers. An "AI" (i.e. LLM) customer service agent/chatbot is just going to piss off customers.
> The Takeaway: Skills are the right abstraction. They formalize the “scripting”-based agent model, which is more robust and flexible than the rigid, API-like model that MCP represents.
Just to not confuse, MCP is like an api but the underlying api can execute an Skill. So, its not MCP vs Skill as a contest. It's just the broad concept of a "flexible" skill vs "parameter" based Api. And again parameter based APIs can also be flexible depending on how we write it except that it lacks SKILL.md in case of Skills which guides llm to be more generic than a pure API.
By the way, if you are a Mac user, you can execute Skills locally via OpenSkills[1] that I have created using apple contianers.
Nothing crazy about it, judging by how much CPU and memory it uses. Now, if it managed to grow features without bringing my M4 Mac with 64GB of ram to a crawl... that's be magic.
sorry I'm not actively working on 10 at the time, but they are in memory and kept open for when I continue working on them. I'm only actively using 2 or 3 at the same time.
I use claude code every day, and havent had a chance to dig super deep into skills, but even though ive read a lot of people describe them and say they're the best thing so far, I still dont get them. Theyre things the agent chooses to call right? They have different permissions? is it a tool call with different permissions and more context? I have yet to see a single post give an actual real-world concrete example of how theyre supposed to be used or a compare and contrast with other approaches.
The prerequisite thought here is that you're using CC to invoke CLI tools.
So now you need to get CC to understand _how_ to do that for various tools in a way that's context efficient, because otherwise you're relying on either potentially outdated knowledge that Claude has built in (leading to errors b/c CC doesn't know about recent versions) or chucking the entirety of a man page into your default context (inefficent).
What the Skill files do is then separate the when from the how.
Consider the git cli.
The skill file has a couple of sentences on when to use the git cli and then a much longer section on how it's supposed to be used, and the "how" section isn't loaded until you actually need it.
I've got skills for stuff like invoking the native screenshot CLI tool on the Mac, for calling a custom shell script that uses the github API to download and pull in screenshots from issues (b/c the cli doesn't know how to do this), for accessing separate APIs for data, etc.
After CC used that skill and it is now in the context, how do you get rid of it later when you don’t need the skill anymore and don’t want to have your context stuffed with useless skill descriptions?
What I find works best for complex things is having one session generate the plan and then dispatching new sessions for each step to prevent context-rot. Not "parallel agents" but "sequential agents."
I think if it literally as a collection of .md files and scripts to help perform some set of actions. I'm excited for it not really as a "new thing" (as mentioned in the post) but as effectively an endorsement for this pattern of agent-data interaction.
So if youre building your own agent, this would be a directory of markdown documents with headers that you tell the agent to scan so that its aware of them, and then if it thinks they could be useful it can choose to read all the instructions into its context? Is it any more than that?
I guess I dont understand how this isnt just RAG with an index you make the agent aware of?
It also looks a lot like a tool that has a description mentioning it has a more detailed MD file the LLM can read for instructions on complex workflows, doesn’t it? MCP has the concept of resources for this sort of thing. I don’t see any difference between calling a tool and calling a CLI otherwise.
I mean it is technically RAG as the LLM is deciding to retrieve a document. But it’s very constrained.
The skills that I use all direct a next action and how to do it. Most of them instruct to use Tasks to isolate context. Some of them provide abstraction specific context (when working with framework code, find all consumers before making changes. add integration tests for the desired state if it’s missing, then run tests to see…) and others just inject only the correct company specific approach to solving only this problem into Task context.
They are composable and you can build the logic table of when an instance is “skilled” enough. I found them worse than hooks with subagents when I started, but now I see them as the coolest thing in Claude code.
The last benefit is nobody on your team even had to know they exist. You can just have them as part of onboarding and everyone can take advantage of what you’ve learned even when working on greenfield projects that don’t have a CLAUDE.md.
Does anyone else struggle with getting Claude to follow even the simplest commands in CLAUDE.md? I’ve completely give up on maintaining it because there’s a coin flip chance it disregards even the simplest instructions. For instance, (after becoming increasingly exasperated and whittling it down over and over), my CLAUDE file now has a single instruction, which says that any ai generated one-off script should be named <foo>.aigen.ts, and I can’t get Claude to follow something as simple as that! It does sometimes, but half the time it disregards the instructions.
I use Claude all the time, and this is probably my biggest issue with it. I just workaround by manually supplying context in prompts, but it’s kind of annoying to do so.
Does anyone else struggle with this or am I just doing something horribly wrong?
I have quite long CLAUDE.md file and claude code follows it almost all the time. When it was ignoring some instructions, i told it to update CLAUDE.md to make sure it does. It emphasized it with uppercase and IMPORTANT! ALWAYS DO ... and that make it work for me like 95% of the time. And I do this for multiple project with different CLAUDE.mds with similar results. I don't know why your experience differs so much.
Definitely experience this. After losing over 2 hours fiddling between a few sessions and still having no actual control I’ve settled on a handful of python and perl/regex scripts to make sure things are following my own conventions
Blog posts like this would really benefit from specific examples. While I can get some mileage out of these tools for greenfield projects, I'm actually shocked that this has proven useful with projects of any substantial size or complexity. I'm very curious to understand the context where such tools are paying off.
It seems to be relative to skill level. If you're less-experienced, you're letting these things write most if not all of your code. If you're more experienced, that's inverted (you write most of the code and let the AI safely pepper things in).
Had a newer employee rewriting some functions with LLMs, but not really admitting to it. I don't really care about the LLM aspect and think it can be quite useful with learning, but I would like to see newer people learning the system before unleashing LLMs entirely on it if that makes any sense. The same developer had a pretty hard time getting an if statement correct that simply checked the length of a string. Lots of mentoring ahead...
But I dunno. I kind of wonder how I would have acted with tech like this available when I first started years ago. As a young engineer and even now I live and breathe the code, obsessing over testing and all the thing that make software engineering so fun. But I can't say whether or not I would have overdepended on an LLM as a young engineer.
It's all kind of depressing IMO, but it is what it is at this point.
Don't rule out laziness! I'm a very experienced senior dev (full stack, embedded, Rust, Python, web everything, etc)... Could I have spent a ton of time learning the ins and outs of Yjs (and the very special way in which you can integrate it with TipRap/prosemirror) in order to implement a concise, collaborative editor? Sure.
Or I could just tell Claude Code to do it and then spend some time cleaning it up afterwards. I had that thing working quite robustly in days! D A Y S!
(Then I had the bright idea of implementing a "track changes" mode which I'm still working on like a week and a half later, haha)
Even if you were already familiar with all that stuff, it's a lot of code to write to make it work! The stylesheets alone... Ugh! So glad I could tell the AI something like, "make sure it implements light and dark mode using VueUse's `useDark()` feature."
Almost all of my "cleanup" work was just telling it about CSS classes it missed when adding dark mode variants. In fact, most of my prompts are asking it to add features (why not?) or cleaning up the code (e.g. divide things into smaller, more concise files—all the LLMs really love to make big .vue files).
"Writing most of the code"? No. Telling it how to write the code with a robust architecture, using knowledge developed over two decades of coding experience: Yes.
I have to reject some things because they'd introduce security vulnerabilities but for the most part I'm satisfied with Claude Code spits out. GPT5, on the other hand... Everything needs careful inspection.
Makes sense. I work for a growth stage startup and most of these apply to our internal mono repo so hard to share specifics. We use this for both new and legacy code each with their own unique AI coding challenges.
If theres enough interest, I might replicate some examples in an open source project.
I find the "slop" thing interesting because—to me—it looks like laziness. In the same way that anyone can tell ChatGPT to write something for them instead of writing it themselves and just having it check the work... Or going through multiple revisions before you're satisfied (with what it wrote).
Code is no different! You can tell an AI model to write something for you and that's fine! Except you have to review it! If the code is bad quality just take a moment to tell the AI to fix it!
Like, how hard is it to tell the AI that the code it just wrote is too terse and hard to read? Come on, folks! Take that extra moment! I mean, I'm pretty lazy when working on my hobby projects but even I'm going to get irritated if the code is a gigantic mess.
Just tell it, "this code is a gigantic mess. Refactor it into concise, human-readable files using a logical structure and make sure to add plenty of explanatory comments where anything might be non-obvious. Make sure that the code passes all the tests when you're done."
> Generally my goal is to “shoot and forget”—to delegate, set the context, and let it work. Judging the tool by the final PR and not how it gets there.
This feels like a false economy to me for real sized changes, but maybe I’m just a weak code reviewer. For code I really don’t care about, I’m happy to do this, but if I ever need to understand that code I have an uphill battle. OTOH reading intermediate diffs and treating the process like actual pair programming has worked well for me, left me with changes I’m happy with, and codebases I understand well enough to debug.
I treat everything I find in code review as something to integrate into the prompts. Eventually, on a given project, you end up getting correct PRs without manual intervention. That's what they mean. You still have to review your code of course!
No, I think it is normal. If it were easy to gain a mental model of the code simply by reading, then debugging would be trivial. The whole point of debugging is that there are differences between your mental model of the code and what the code is actually doing, that sometimes can't be uncovered unless you step through it line by line even if you're the one who wrote it.
It is why I am a bit puzzled by the people who use an LLM to generate code in anything other than a "tightly scoped" fashion (boilerplate, throwaway code, standalone script, single file, or at the function level). I'm not sure how that makes your job later on any easier if you have even a worse mental model of the code because you didn't even write it. And debugging is almost usually more tedious than writing code, so you've traded off the fun/easy part for a more difficult one. Seems like a faustian deal.
Skills are also a convenient way for writing self-documenting packages. They solve the problem of teaching the LLM how to use a library.
I have started experimenting with a skills/ directory in my open source software, and then made a plugin marketplace that just pulls them in. It works well, but I don't know how scalable it will be.
I've been having good results with Roo Code/Cline + Claude Code as the provider. I feel like using CC CLI effectively requires manually recreating Roo/Cline
I really enjoyed reading this. One thought I had on the issue of paths in Claude.md
My concern with hardcoding paths inside a doc, it will likely become outdated as the codebase evolves.
One solution would be to script it and have it run pre commit to regenerate the Claude.md with the new paths.
There probably is potential for even more dev tooling that 1. Ensure reference paths are always correct, 2. Enforces standard for how references are documented in Claude.md (and lints things like length)
Perhaps using some kind of inline documentation standard like jsdoc if it’s a ts file or a naming convention if it’s an Md file
Example:
// @claude.md
// For complex … usage or if you encounter a FooBarError, see ${path} for advanced troubleshooting steps
I feel like these posts are interesting, but become irrelevant quickly. Does anyone actually follow these as guides, or just consume them as feedback for how we wish we could interface with LLMs and the workarounds we currently use?
Right now these are reading like a guide to prolog in the 1980s.
Given that this space is so rapidly evolving, these kinds of posts are helpful just to make sure you aren't missing anything big. I've caught myself doing something the hard way after reading one of these. In this case, the framing is basically man pages for CLIs was a helpful description of sills that gives me some ideas about how to improve interaction with an in-house CLI my co. uses.
Yeah I like to think not everyone can spend their day exploring/tinkering with all these features so it's handy to just snapshot what exists and what works/doesn't.
I wouldn't say I follow them as guides, but I think the field is changing quickly enough that it's good, or at least interesting, to read what's working well for other people.
This one is already out of date. The bit on the top about allocating space in CLAUDE.md for each tool is largely a waste of tokens these days. Use the skills feature.
The skill lets you compress the amount loaded to just the briefest description, with the “where do I go to get more info” being implicit. You should use a SKILL.md for evry added tool. At which point, putting instructions in CLAIDE.md becomes redundant and confusing to the LLM.
Are "agents" really as useful as the hype makes them out to be?
Most of the time I'm just pasting code blocks directly into raycast and once I've fixed the bug or got the properly transformed code in the shape that I aimed for, then I paste it back into neovim. Next i'm going to try out "opencode"[0], because I've heard some good things about it. For now, I'm happy with my current workflow.
Yes, agents are really as useful as the hype makes them out to be. Certainly they’re more useful on bigger code bases but even small ones the agent provides more value than you may expect. I have a small side project probably around 10k loc and the ai agents still plow through tokens. End of the day letting the ai read your entire code base really does let it make more intelligent design decisions. Many other benefits as well. Iteration is much faster without copy paste
I don't understand how people use the `git worktree` workflow. I get that you want to isolate your work, but how do you deal with dev servers, port conflicts and npm installs? When I tried it, it was way more hassle than it was worth.
I generally like to use it. But I one project in the org which simply can’t work because the internal built system expects a normal .git directory at the root.
Means I have to rewrite some of the build code that isn’t aware of this git feature. And yes we use a library to read from git but not the git cli or a more recent compatible one that understands that the current work tree is not the main one.
Agree, depending on the repo and changes it’s hard with local dev servers. It sometimes works well if you don’t need local dockers and want to outsource git workflow to CC as well. Then it can do on that branch whatever it wants and main work is in another worktree with more steering and or docker env.
I have a bash script that creates the worktree, copies env over and changes the ports of containers and the services. I then can proxy the "real" port to any worktree, it's common I'll have 3 worktrees active to switch back and forth
Hooks are underutilized and will be critical for long-running and better agent performance. Excited to release Cupcake in the coming weeks. What I started building when i made the feature request for hooks.
"It shifts us from a “per-seat” license to “usage-based” pricing, which is a much better model for how we work. It accounts for the massive variance in developer usage (We’ve seen 1:100x differences between engineers)."
This is how we are doing it too, for the same reasons. For now, much easier to administer than trying to figure out who spends enough to get a real Claude Code seat. The other nice thing about using API keys is that you basically never hit rate limits.
Does anyone have any suggestions on making Claude prefer to use project internal abstractions and utility functions? My C++ project has a lot of them. If I just say something like "for I/O and networking code, check IOUtils.h for helpers" then it often doesn't do that. But mentioning all helper functions and classes in the context also seems like a bad idea. What's the best way? Are the new Skills a solution?
Skills seem like the way forward, but Claude still needs to be convinced to activate the skill. If that's not happening reliably, hooks should be able to help.
A sibling comment on hooks mentions some approaches. You could also try leveraging the UserPromptSubmit hook to do some prompt analysis and force relevant skill activation.
I wonder how well a sentence or two in CLAUDE.md, saying to search the local project for examples of similar use cases or use of internal libraries, would work.
Hooks can also be useful for this. If it's using the wrong APIs then can hint on write or block on commit with some lint function that checks for this.
What type of projects are you guys building? I bought Max and these features to try it out to build a more complex project (ROS2) and that does not seem to work out at all… HTML page, yes, embedded project not so much.
With all the LLM coding assistants, you have to get a feel for each model and which extension/interface you're using with them. Not only that, but it's also dependent on your project!
For example, if you're writing a command line tool in Python, it doesn't really matter what model you use since they're all really great at Python (LOL). However, if you're writing a complicated SPA that uses say, Vue 3 with Vite (and some fancy CSS framework) and Python w/FastAPI... You want the "smartest" model that knows about all these things at once (and regularly gets updated knowledge of the latest versions of things). For me, that means Claude Code.
I am cheap though and only pay Anthropic $20/month. This means I run out of Claude Credits every damned week (haha). To work around this problem, I used to use OpenAI's pay-per-use API with gpt5-mini with VS Code's Copilot extension, switching to GPT5-codex (medium) with the Codex extension for more complicated tasks.
Now that I've got more experience, I've figured out that GPT5-codex costs way too much (in API credits) for what you get in nearly all situations. Seriously: Why TF does it use that much "usage". Anyway...
I've tried them all with my very, very complicated collaborative editor (CRDTs), specifically to learn how to better use AI coding assistants. So here's what I do now:
* Ollama cloud for gpt-oss:120b (it's so fast!)
* Claude Code for everything else.
I cannot understate how impressed I am with gpt-oss:120b... It's like 10x faster than gpt5-mini and yet seems to perform just as well. Maybe better, actually because it forces you to narrow your prompts (due to smaller context window). But because it's just so damned fast, that doesn't matter.
With Claude Code, it's like magic: You give it a really f'ing complicated thing to troubleshoot or implement and it just goes—and keeps going until it finishes or you run out of tokens! It's a, "the future is now!" experience for sure.
With gpt-oss:120b it's more like having an actual conversation, where the only time you stop typing is when you're reviewing what it did (which you have to do for all the models... Some more than others).
FYI: The worst is Gemini 2.5. I wouldn't even bother! It's such trash, I can't even fathom how Google is trying to pass it off as anything more than a toy. When it decides to actually run (as opposed to responding with, "Failed... Try again"), it'll either hallucinate things that have absolutely nothing to do with your prompt or it'll behave like some petulant middle school kid that pretend to spend a lot of time thinking about something but ultimately does nothing at all.
They said don’t __just__ use never, and explain that as meaning you should give an alternative with every “never” clause, which is what their example does :)
Are there any of those CLI clients (coded in plain and simple C, or basic python/perl without 1 billion of expensive dependencies) able to access those 'coding AI' prompt anonymously then rate limited?
If no anonymous access is provided, is there a way to create an account with a noscript/basic (x)html/classic web browsers in order to get an API key secret?
Because I do not use web engines from the "whatng" cartel.
To add insult to injury, my email is self-hosted with IP literals to avoid funding the DNS people which are mostly now in strong partnership with the "whatng" cartel (email with IP literals are "stronger" than SPF since it does the same and more). An email is often required for account registration.
Just my curiosity: Why are you producing so much code? Is it because it is now possible to do so with AI, or because you have a genuine need (solid business usecase) that requires a lot of code?
I just started developing self-hosted services largely with AI.
It wasn't possible before for me to do any of this at this kind of scale. Before, getting stuck on a bug could mean hours, days, or maybe even weeks of debugging. I never made the kind of progress I wanted before.
Many of the things I want, do already exist, but are often older, not as efficient or flexible as they could be, or just plain _look_ dated.
But now I can pump out react/shadcn frontends easily, generate apis, and get going relatively quickly. It's still not pure magic. I'm still hitting issues and such, but they are not these demotivating, project-ending, roadblocks anymore.
I can now move at a speed that matches the ideas I have.
I am giving up something to achieve that, by allowing AI to take control so much, but it's a trade that seems worth it.
Often code in SaaS companies like ours is indeed how we solve customer problems. It's not so much the amount of code but the rate (code per time) we can effectively use to solve problems/build solutions. AI, when tuned correctly, lets us do this faster than ever possible before.
If you don't want to think and offload it to llm they burn through a lot of tokens to implement in a non-efficient way something you could often do in 10 lines if you though about it for a few minutes.
I’ve just implemented a proof of concept that involved an API, a MCP server, an Authorization Server, a React frontend, token validation and proof of possession on the client, a CIBA flow for authentication… took a week , and I don’t even know the technologies used very well, it was all TypeScript but I work on JVM languages normally. This was a one off for a customer and I was able to show a fairly complex workflow end to end and what each part involves. I let the LLM write most of it but I understand every line and did have to make manual adjustments (though to be honest, I could easily explain to the LLM what I needed changed and given my experience it would eventually get there.
If you tell me I didn’t really need a LLM to be able to do all that in a week and just some thought and 10 lines of code would do, I suspect you are not really familiar with the latest developments in AI and just vastly underestimates the capabilities they have to do tricky stuff.
In a large project with decent code structure there can be quite a bit of boilerplate, convention, testing required. Also we are not talking about a 10-line change. More like 10k line feature.
Before LLMs we simply wouldn't implement many of those features since they were not exactly critical and required a lot of time, but now when the required development time is cut signifficantly, they suddenly make sense to implement.
If you have no modifications or customization of Claude code then it comes down to a preference for proactivity (codex) or a bit more restraint.
If you are using literally any of Claude Code’s features the experience isn’t close, and regardless of model preference (Claude is my least favorite model by far) you should probably use Claude code. It’s just a much more extensible product for teams.
Which features are preferable to higher quality output?
Losing access to GPT 5 Pro is also a big hit… it is by far the best for reading full files/repos and creating plans (though it also by far has the worst out of the box tooling)
CC has better agent tools and is faster. The ability to switch from plan mode to execution mode and back is huge. Toggling thinking also. And of course they are innovating all of these agentic features like MCP, sub-agents, skills, etc...
Codex writes higher quality code, but is slower and less feature rich. I imagine this will change within months. The jury is still out. Exciting times!
I guess I don’t understand wanting faster and worse for much work, and some of the features like subagents are dubious or like skills and planning mode are minor conveniences over skill files mentioned by agents.md and toggling read only mode or using a plan file. After all those latter features are just conveniences for assembling context.
Maybe CC users haven’t figured out how to parallelize their work because it’s fast enough to just wait or be distracted, and so the Codex waiting seems unbearable.
I use both at the same time. CC seems to have better access to web and researching capabilities compared to Codex. Maybe I'm not using Codex right or missing something, but it has frequent troubles browsing internet. Also Claude Code is faster. So I use it when I know it can handle the task.
When touch typing and talking to someone, I accidentally typed something to claude with my fingers off the home row, e.g. ttoubg kuje tgus ubti tge ckayde cide ternubak. Claude understood it just fine. Didn't even remark on it.
It truly is an idiot savant. It's absurdly good, and then if there is any unaccounted complexity, ugh let's just make the tests stubs.... tests pass now.
wait until you start conversing with it. It's been a game changer for me how I use Claude CLI. It suits my workflow fine since my sessions are intense in focus I have to bring and I iterate with it; I just haven't found _a way_ where I can give it a large thing to work on and that it will not deviate. I do one focused thing at a time, review, test, alter code and then repeat. With voice mode it's been great since I can talk with it while walking fast on a treadmill. It's bizarre, star trekish, and it works. I wish I could have a stop word with whisper, since I do tend to think long between sentences in this mode, and I wish I could stop it with voice while it's talking, but I found a flow that it doesn't matter that much.
No thanks. I rather write the code myself that use generated slop. I actually like to code and see little benefit in other peoples copypaste code (thats essentially what ai slop is really)
I feel or have the fear that the world will tumble and crack under the sheer amount of code we produce and can’t be maintained because at one point no one human can understand all the stuff that was written.
At the moment though I also code on and off with an agent. I’m not ready or willing to only vibe code my projects. For one is the fact that I had tons of examples where the agent gaslighted me only to turn around at the last stage. And in some cases the code output was to result focused and didn’t think about the broader general usage. And sure that’s in part because I hold it wrong. Don’t specify 10million markdown files etc. But it’s a feedback loop system. If I don’t trust the results I don’t jump in deeper. And I feel a lot of developers have no issue with jumping ever deeper. Write MCPs now CLIs and describe projects with custom markdown files.
But I think we really need both camps. Otherwise we don’t move forward.
> I feel or have the fear that the world will tumble and crack under the sheer amount of code we produce and can’t be maintained because at one point no one human can understand all the stuff that was written.
IMO the best advice in life is try not to be fearful of things that happen to everyone and you can't change.
Good news! What you are afraid of will happen, but it'll happen to everyone all at once, and nothing you can do can change it.
So you no longer need to feel fear. You can skip right on over to resignation. (We have cookies, for we are cooked)
> Finally, we keep this file synced with an AGENTS.md file to maintain compatibility with other AI IDEs that our engineers might be using.
I researched this the other day, the recommended (by Anthropic) way to do this is to have a CLAUDE.md with a single line in it:
Then keep your actual content in the other file: https://docs.claude.com/en/docs/claude-code/claude-code-on-t...This is one thing I think they need to get in-line with, and rename CLAUDE.md to AGENTS.md to follow convention.
To be fair, I think Anthropic/Claude started doing CLAUDE.md before AGENTS.md was a thing.
I'd say if we criticize that they did not think that far ahead, we are still pretty fair.
I strongly disagree. It is extremely annoying when I cannot make different agents read a different prompt file when I needed to tell them different things. Cursor currently autoloads AGENTS.md without a way to disable it and it sucks.
They ain't giving up that free marketing.
We have an AGENTS.md symlinked to CLAUDE.md, seems to work fine.
This is the way.
Yeah that's probably a slightly cleaner way of doing it.
You think it would be a good idea to use a symlink instead?
I use symbolic links, and Claude Code often gets confused, requiring several iterations to understand that the CLAUDE.md file is actually a symbolic link to AGENTS.md, and that these are not two different, duplicate files
The recommended approach has the advantage of separating information specific to Claude Code, but I think that in the long run, Anthropic will have to adopt the AGENTS.md format
Also, when using separate files, memories will be written to CLAUDE.md, and periodic triaging will be required: deciding what to leave there and what to move to AGENTS.md
I'm still not 100% sure I understand what a symlink in a git repository actually does, especially across different operating systems. Maybe it's fine?
Anthropic say "put @AGENTS.md in your CLAUDE.md" file and my own experiments confirmed that this dumps the content into the system prompt in the same way as if you had copied it to CLAUDE.md manually, so I'm happy with that solution - at least until Anthropic give in and support AGENTS.md directly.
It just creates the same symlink on any other checkout. (On Linux/macOS at least, Windows I believe requires local settings changes.)
Only sane (guaranteed portable) option is for it to be a relative symlink to another file within the same repo, of course. i.e. CLAUDE.md would be -> 'AGENTS.md', not '/home/simonw/projects/pelicans-on-bicycles/AGENTS.md' or whatever.
On windows, it depends on the local git configuration. It’s not something I’ve been happy with, especially since symlinks also behave differently again when you’re running a docker container to get your windows usable for development.
I have AGENTS.md symlinked to CLAUDE.md and it works fine in my repos.
But I can’t speak to it working across OS.
Confirm on a new clone that if you modify a file that the other is updated.
I thought git by default treats symlinks simply as file copies when cloning new.
Ie git may not be aware of the symlink.
git very much supports symlinks. Although depending on the system config it might not create actual symlinks on Windows.
In my experience, neither Claude nor any other agent actually reads AGENTS.md (or CLAUDE.md or anything else) without being told to explicitly every session.
I've sniffed Claude Code's HTTP traffic and confirmed that the CLAUDE.md file content (and AGENTS.md if it is @-referenced) is automatically included in the system prompt without it having to perform any additional file read operations.
Wow! That's....
discouraging, actually, considering how frequently Claude ignores my AGENTS.md guidance.
Did you notice the @-referencing requirement? AGENTS.md is not included by default, but CLAUDE.md should be
You're going to immediately be overwhelmed with code, comments, markdown files, and all manner of extra text that may or may not solve your problem. The chore then becomes PR reviews of tons of iffy code, and if you merge too much of it, after too long your codebase will be a bloated mess that even the craftiest prompts won't untangle.
It's best to use to brainstorm and implement very targeted prompts, set and forget is crazy.
Completely agree. The quoted strategy only works for small hobbyist codebases.
I really like this take on MCP: https://blog.sshh.io/i/177742847/mcp-model-context-protocol
> Instead of a bloated API, an MCP should be a simple, secure gateway that provides a few powerful, high-level tools [...] In this model, MCP’s job isn’t to abstract reality for the agent; its job is to manage the auth, networking, and security boundaries and then get out of the way.
Thanks! I def don't think I would have guessed this use case when MCP first came out, but more and more it seems Claude just yearns for scripting on data rather than a bunch of "tools". My/MCPs job has become just getting it that data.
Have you tried using light CLIs rather than MCP? I’ve found that CLIs are just easier for Claude, especially if you write them with Claude and during planning instruct it to think about adding guidance to users who get confused.
Our auth, log diving, infra state, etc, is all usable via cli, and it feels pretty good when pointing Claude at it.
Yeah if that's possible or you are willing to build it, that's the right solution. Today pretty much all of my integrations are pure CLIs like that rather than MCPs.
You can do anything you want via a CLI but MCP still exists as a standard that folks and platforms might want to adopt as a common interface.
MCPs still lack scoping in most cases.
It's not that practical to have an MCP that can connect to, for example, ALL of your corporate Google Drive. Not happening.
Why isn't it possible to limit it to a specific whitelisted set?
> Why isn't it possible to limit it to a specific whitelisted set?
It is but you need to use an MCP client and servers that implement Roots.
Agreed. My only MCP is a code interpreter. I also recently started experimenting with making an MCP “proxy” which acts a better harness that lets the agent call MCP from within a code interpreter [1]
But in general I still don’t really use MCP. Agents are just so good at solving problems themselves. I wish MCP would mostly focus at the auth part instead of the tool part. Getting an agent access to an API with credentials usually gives them enough power to solve problems on their own.
[1]: https://x.com/mitsuhiko/status/1984756813850374578?s=46
This is how MCP works if you use it for as essential an internal tool API gateway (stateless http) instead of a client facing service that end users are connecting directly to. It's basically just OpenAPI but slightly more tuned for LLM inference.
Don't sleep on using Claude Code to improve your Claude Code config. Switch to plan mode and try the following prompt:
read the document at https://blog.sshh.io/p/how-i-use-every-claude-code-feature and tell me how to improve my Claude code setup
"Claude Code isn’t just an interactive CLI; it’s also a powerful SDK for building entirely new agents—..."
Em dash and "it's not X, it's Y" in one sentence. Tired of reading posts written by AI. Feels disrespectful to your readers
I have the same instinctive response to reading AI generated stuff, but I'm coming to a more moderate position where I'm trying to judge the content on the content itself. For example, in a post like this, it doesn't bother me at all because it's still an extremely useful reference, and the author clearly read through, organized, and edited the output. This is a good example of usage of AI in my opinion.
The people who just copy paste output from ai and ship it as a blog post however, deserve significant condemnation for that.
Writing is a tool for thought I feel, and when you're outsourcing your thought, it detracts from whatever you intended to say. I guess if it was that heavily-edited such a telltale sign wouldn't have remained.
I use AI for code, but I never use it for any writing that is for human eyes.
Not worried about hallucinations?
I definitely would be, though this part I consider mitigation:
> the author clearly read through, organized, and edited the output.
Also worth noting, I've read plenty of human written stuff that has errors in it, so I read everything skeptically anyway.
If the author wrote the draft then reread it throughly then it most likely would only have human induced hallucinations
The internet is dead. Long live the internet.
the eternal september 2: it's eternaler this time
We're on Eternal September 3 now.
The iPhone brought us Eternal September 2.
> Tired of reading posts written by AI.
Didn’t realize you were forced to read this?
> Feels disrespectful to your readers
I didn’t feel disrespected—I felt so respected I read the whole thing.
Nothing there implies I was forced. I read the whole thing. It was not disclosed that it was written (in part?) by AI. And That sentence I quoted was pretty far along.
> It was not disclosed that it was written (in part?) by AI
Who gives a shit?
If you can’t stand AI writing and you made it pretty far along before getting upset, who are you mad at, the author or yourself? Would you be happier if you found out this was written without AI and that you were just bad at detecting AI writing?
> /clear + /catchup (Simple Restart): My default reboot. I /clear the state, then run a custom /catchup command to make Claude read all changed files in my git branch.
I've found myself doing similar workarounds. I'm guessing anthropic will just make the /compact command do this instead soon enough.
I've found the latency of /compact makes it unusable. Perhaps this is just the result of my waiting until I have 0% context remaining.
Fun fact, a large chunk of context is reserved for compaction. When you are shown that you have "0% context remaining," it's actually like 30% remaining that's reserved for compaction.
And yet, for some reason I feel like 50% of the time, compaction fails because it runs out of context or hits (non-rate) API limits.
Weirdly, I’ve found that when that happens I can close Claude and then run `claude --continue` and now it has room to compact. Makes no sense.
But I have no idea what state it will be in after compact, so it’s better to ask it to write a complete and thorough report including what source files to read. Lot more work but better than going off the rails.
The old rule of thumb is that when you compact, you've already lost.
Kinda sad if 3000 words is now considered "too long to read through rather use as reference" but some interesting points, I'd be keen to see an even longer version with actual examples instead of placeholder ones.
Yeah I'm fairy pessimistic about how much folks will read
100%. I was excited when i read that disclaimer and found myself disappointed by the limited content. That said, i did get a couple tidbits out of it.
> If you’re not already using a CLI-based agent like Claude Code or Codex CLI, you probably should be.
Are the CLI-based agents better (much better?) than the Cursor app? Why?
I like how easy it is to get Cursor to focus a particular piece of code. I select the text and Cmd-L, saying "fix this part, it's broken like this ____."
I haven't really tried a CLI agent; sending snippets of code by CLI sounds really annoying. "Fix login.ts lines 148-160, it's broken like this ___"
Better? Hard to say. Different? Yes. Worth evaluating? Absolutely. Using it for 30 minutes will answer your question better than any reply here. I think you'll answer your own question quickly.
I've been coding seriously for about 15 years. No single tool has changed how I code more than claude code and I'm including non-"AI" tooling/services. This sounds like I'm shilling but I am not affiliated. It's played a large part in injecting my passion back into building stuff.
Yeah I started with Cursor, went hybrid, and then in the last month or so I've totally swapped over.
Part of it is the snappy more minimal UX but also just pure efficacy seems consistently better. Claude does its best work in CC. I'm sure the same is true of Codex.
Cursor Composer appears to have this type of coupling and uses IDE resources better than other models on average.
Seems you haven't heard of Cursor 2.0
https://cursor.com/blog/2-0
Claude is able to detect the lines of code selected in vscode anyway
As-is Gemini CLI and Codex. I run my CLIs in VSC and only using it as a file browser.
They all have optional ide integration, e.g Claude knows the active vscode tab and highlighted lines.
Is that better than Cursor? Same? Just different?
All I can say is when I switched from Cursor to Claude it took me less than 24 hours to realise I wouldn’t go back. The extra UI Cursor slaps on to VS Code is just bloat, which I found quite buggy (might be better now though), and the output was nowhere near as good. Maybe things have improved since I switched but Claude CLI with VS Code is giving me no reasons to want to try anything else. Cursor seemed like a promising and impressive toy, Claude CLI is just a great product that’s delivering value for me every day.
Vscode has agents built in now, have you used that UI?
That particular part is the same, roughly. The bigger issue is just that CC's a better agent than Cursor, last I checked.
There's even an official Anthropic VS Code extension to run CC in VS Code. The biggest advantage is being able to use VS Code's diff views, which I like more than in the terminal. But the VS Code CC extension doesn't support all the latest features of the terminal CC, so I'm usually still in the terminal.
Yes and you can select multiple files to give it focus. It can run anything in your PATH too. Eg it's pretty good at using `gh` and so on
Direct use of Codex + GPT5 or Claude Code CLI gives a better result, compared to using the same models in Cursor. I've compared both. Cursor applies some of their augmentation, which reduces the output size, probably to save on tokens.
Claude is just better at coding than cursor.
Really, the interface isn't a meaningful part of it. I also like cmd-L, but claude just does better at writing code.
...also, it's nice that Anthropic is just focusing on making cool stuff (like skills), while the folk from cursor are... I dunno. Whatever it is they're doing with cursor 2.0 :shrug:
Cursor can use the Claude Sonnet and Claude Opus LLMs, so I would expect output to be quite similar in that respect.
The agentic part of the equation is improving on both sides all the time.
There's something in the prompting, tooling, heuristics inside the Claude Code CLI itself that makes it more than just the model it's talking to and that becomes clear if you point your ANTROPHIC_URL at another model. The results are often almost equivalent.
Whereas I tried Kilo Code and CoPilot and JetBrain's agent and others direct against Sonnet 4 and the output was ... not good ... in comparison.
I have my criticisms of Claude but still find it very impressive.
Claude Code is much more efficient even compared to Cursor using the Anthropic models. The planning and tool use is much better.
I use Claude and Codex in VS Code and they work really well.
> I select the text and Cmd-L, saying "fix this part, it's broken like this
This flow works well.
Man, I just want to point out how different programming has become in 2 years.
We are literally having to play mind games with our tooling now to convince and coax it into doing certain things. I swear we are not far off from having to call in sick to work because your LLM "is going through a lot right now and just needs a mental health break".
I think the final straw for me was when I saw people saying they were getting better results by writing some parts of the prompts in all caps.
I just don't wanna deal with any of this.
As fascinating as these tools can be - are we (the industry) once again finding something other than our “customer” to focus our brains on (see Paul Graham’s “Top idea in your mind” essay)?
It seems so ... LLM-based coding tools are mostly about speed and cost of development - corporate accounting metrics, but what customers care about is mostly product features (& lack of bugs).
There is no customer advantage to developing cheap and fast if the delivered product isn't well conceived from a current and future customer-needs perspective, and a quickly shipped product full of bugs isn't going to help anyone.
I think the same goes for AI in general - CEOs are salivating over adopting "AI" (which people like Altman and Amodei are telling them will be human level tomorrow, or yesterday in the case of Amodei), and using it to reduce employee head count, but the technology is nowhere near the human level needed to actually benefit customers. An "AI" (i.e. LLM) customer service agent/chatbot is just going to piss off customers.
What do you mean by "customer"? Because I'm also using these tools to understand the customer better.
> The Takeaway: Skills are the right abstraction. They formalize the “scripting”-based agent model, which is more robust and flexible than the rigid, API-like model that MCP represents.
Just to not confuse, MCP is like an api but the underlying api can execute an Skill. So, its not MCP vs Skill as a contest. It's just the broad concept of a "flexible" skill vs "parameter" based Api. And again parameter based APIs can also be flexible depending on how we write it except that it lacks SKILL.md in case of Skills which guides llm to be more generic than a pure API.
By the way, if you are a Mac user, you can execute Skills locally via OpenSkills[1] that I have created using apple contianers.
1. OpenSkills -https://github.com/BandarLabs/open-skills
Crazy how fast Claude Code is evolving, every week there’s something new to learn, and it just keeps getting better.
Nothing crazy about it, judging by how much CPU and memory it uses. Now, if it managed to grow features without bringing my M4 Mac with 64GB of ram to a crawl... that's be magic.
My m1 macbook pro works fine with +10 claude code sessions open at the same time (iTerm2). Are you using a terminal with a memory leak perhaps?
I'm using iterm2. Memory leak and cpu problems are very well documented in the github issues.
I have home server (cost around $150) with 16 GB RAM also running Claude Code fine.
How are you managing 10 parallel agents??
I use Windows Terminal. Rename tab.
My current project I have a top level chat , then one chat in each of the four component sub directories.
I have a second terminal with QA-feature
So 10 tabs total . Plus I have one to run occasional commands real quick (like docker ps).
I’m using qwen.
That's a lot of cognitive load to manage especially with how fast CC has become, do you review the output at all?
The users review the output
sorry I'm not actively working on 10 at the time, but they are in memory and kept open for when I continue working on them. I'm only actively using 2 or 3 at the same time.
Huh, Claude Code barely uses any system ressources. Are you sure it's Claude Code and not some Electron app that hasn't been updated for Tahoe?
Claude doesn't do much of anything on the local machine. I run it on a Macbook Air and a piddly 2vCPU 4GB VPS. Works fine.
CC uses very little system resources.
I use claude code every day, and havent had a chance to dig super deep into skills, but even though ive read a lot of people describe them and say they're the best thing so far, I still dont get them. Theyre things the agent chooses to call right? They have different permissions? is it a tool call with different permissions and more context? I have yet to see a single post give an actual real-world concrete example of how theyre supposed to be used or a compare and contrast with other approaches.
The prerequisite thought here is that you're using CC to invoke CLI tools.
So now you need to get CC to understand _how_ to do that for various tools in a way that's context efficient, because otherwise you're relying on either potentially outdated knowledge that Claude has built in (leading to errors b/c CC doesn't know about recent versions) or chucking the entirety of a man page into your default context (inefficent).
What the Skill files do is then separate the when from the how.
Consider the git cli.
The skill file has a couple of sentences on when to use the git cli and then a much longer section on how it's supposed to be used, and the "how" section isn't loaded until you actually need it.
I've got skills for stuff like invoking the native screenshot CLI tool on the Mac, for calling a custom shell script that uses the github API to download and pull in screenshots from issues (b/c the cli doesn't know how to do this), for accessing separate APIs for data, etc.
After CC used that skill and it is now in the context, how do you get rid of it later when you don’t need the skill anymore and don’t want to have your context stuffed with useless skill descriptions?
You'd need to do the "/clear" or other context manipulations.
What I find works best for complex things is having one session generate the plan and then dispatching new sessions for each step to prevent context-rot. Not "parallel agents" but "sequential agents."
Maybe these might be handy: - https://github.com/anthropics/skills - https://www.anthropic.com/engineering/equipping-agents-for-t...
I think if it literally as a collection of .md files and scripts to help perform some set of actions. I'm excited for it not really as a "new thing" (as mentioned in the post) but as effectively an endorsement for this pattern of agent-data interaction.
apparently I missed Simon Willison's article, this at least somewhat explains them: https://simonwillison.net/2025/Oct/16/claude-skills/
So if youre building your own agent, this would be a directory of markdown documents with headers that you tell the agent to scan so that its aware of them, and then if it thinks they could be useful it can choose to read all the instructions into its context? Is it any more than that?
I guess I dont understand how this isnt just RAG with an index you make the agent aware of?
It also looks a lot like a tool that has a description mentioning it has a more detailed MD file the LLM can read for instructions on complex workflows, doesn’t it? MCP has the concept of resources for this sort of thing. I don’t see any difference between calling a tool and calling a CLI otherwise.
I mean it is technically RAG as the LLM is deciding to retrieve a document. But it’s very constrained.
The skills that I use all direct a next action and how to do it. Most of them instruct to use Tasks to isolate context. Some of them provide abstraction specific context (when working with framework code, find all consumers before making changes. add integration tests for the desired state if it’s missing, then run tests to see…) and others just inject only the correct company specific approach to solving only this problem into Task context.
They are composable and you can build the logic table of when an instance is “skilled” enough. I found them worse than hooks with subagents when I started, but now I see them as the coolest thing in Claude code.
The last benefit is nobody on your team even had to know they exist. You can just have them as part of onboarding and everyone can take advantage of what you’ve learned even when working on greenfield projects that don’t have a CLAUDE.md.
Does anyone else struggle with getting Claude to follow even the simplest commands in CLAUDE.md? I’ve completely give up on maintaining it because there’s a coin flip chance it disregards even the simplest instructions. For instance, (after becoming increasingly exasperated and whittling it down over and over), my CLAUDE file now has a single instruction, which says that any ai generated one-off script should be named <foo>.aigen.ts, and I can’t get Claude to follow something as simple as that! It does sometimes, but half the time it disregards the instructions.
I use Claude all the time, and this is probably my biggest issue with it. I just workaround by manually supplying context in prompts, but it’s kind of annoying to do so.
Does anyone else struggle with this or am I just doing something horribly wrong?
I have quite long CLAUDE.md file and claude code follows it almost all the time. When it was ignoring some instructions, i told it to update CLAUDE.md to make sure it does. It emphasized it with uppercase and IMPORTANT! ALWAYS DO ... and that make it work for me like 95% of the time. And I do this for multiple project with different CLAUDE.mds with similar results. I don't know why your experience differs so much.
Definitely experience this. After losing over 2 hours fiddling between a few sessions and still having no actual control I’ve settled on a handful of python and perl/regex scripts to make sure things are following my own conventions
Same experience, although I gave up already a while ago so not sure if CLAUDE.md is better followed in newer versions.
It's a known issue among us Claudemasters.
Blog posts like this would really benefit from specific examples. While I can get some mileage out of these tools for greenfield projects, I'm actually shocked that this has proven useful with projects of any substantial size or complexity. I'm very curious to understand the context where such tools are paying off.
It seems to be relative to skill level. If you're less-experienced, you're letting these things write most if not all of your code. If you're more experienced, that's inverted (you write most of the code and let the AI safely pepper things in).
Had a newer employee rewriting some functions with LLMs, but not really admitting to it. I don't really care about the LLM aspect and think it can be quite useful with learning, but I would like to see newer people learning the system before unleashing LLMs entirely on it if that makes any sense. The same developer had a pretty hard time getting an if statement correct that simply checked the length of a string. Lots of mentoring ahead...
But I dunno. I kind of wonder how I would have acted with tech like this available when I first started years ago. As a young engineer and even now I live and breathe the code, obsessing over testing and all the thing that make software engineering so fun. But I can't say whether or not I would have overdepended on an LLM as a young engineer.
It's all kind of depressing IMO, but it is what it is at this point.
Don't rule out laziness! I'm a very experienced senior dev (full stack, embedded, Rust, Python, web everything, etc)... Could I have spent a ton of time learning the ins and outs of Yjs (and the very special way in which you can integrate it with TipRap/prosemirror) in order to implement a concise, collaborative editor? Sure.
Or I could just tell Claude Code to do it and then spend some time cleaning it up afterwards. I had that thing working quite robustly in days! D A Y S!
(Then I had the bright idea of implementing a "track changes" mode which I'm still working on like a week and a half later, haha)
Even if you were already familiar with all that stuff, it's a lot of code to write to make it work! The stylesheets alone... Ugh! So glad I could tell the AI something like, "make sure it implements light and dark mode using VueUse's `useDark()` feature."
Almost all of my "cleanup" work was just telling it about CSS classes it missed when adding dark mode variants. In fact, most of my prompts are asking it to add features (why not?) or cleaning up the code (e.g. divide things into smaller, more concise files—all the LLMs really love to make big .vue files).
"Writing most of the code"? No. Telling it how to write the code with a robust architecture, using knowledge developed over two decades of coding experience: Yes.
I have to reject some things because they'd introduce security vulnerabilities but for the most part I'm satisfied with Claude Code spits out. GPT5, on the other hand... Everything needs careful inspection.
Makes sense. I work for a growth stage startup and most of these apply to our internal mono repo so hard to share specifics. We use this for both new and legacy code each with their own unique AI coding challenges.
If theres enough interest, I might replicate some examples in an open source project.
Whats interesting to see is not the project setup but the resulted generated code in a mid-sized project.
To see if it is easy to digest, no repeated code etc or is it just slop that should be consumed by another agent and never by human.
I find the "slop" thing interesting because—to me—it looks like laziness. In the same way that anyone can tell ChatGPT to write something for them instead of writing it themselves and just having it check the work... Or going through multiple revisions before you're satisfied (with what it wrote).
Code is no different! You can tell an AI model to write something for you and that's fine! Except you have to review it! If the code is bad quality just take a moment to tell the AI to fix it!
Like, how hard is it to tell the AI that the code it just wrote is too terse and hard to read? Come on, folks! Take that extra moment! I mean, I'm pretty lazy when working on my hobby projects but even I'm going to get irritated if the code is a gigantic mess.
Just tell it, "this code is a gigantic mess. Refactor it into concise, human-readable files using a logical structure and make sure to add plenty of explanatory comments where anything might be non-obvious. Make sure that the code passes all the tests when you're done."
It's always laziness. The people that do the bare minimum will likely continue to do so, regardless of AI.
I think we'll be dealing with slop issues for quite some time, but I also have hopes that AI will raise the bar of code in general.
> Generally my goal is to “shoot and forget”—to delegate, set the context, and let it work. Judging the tool by the final PR and not how it gets there.
This feels like a false economy to me for real sized changes, but maybe I’m just a weak code reviewer. For code I really don’t care about, I’m happy to do this, but if I ever need to understand that code I have an uphill battle. OTOH reading intermediate diffs and treating the process like actual pair programming has worked well for me, left me with changes I’m happy with, and codebases I understand well enough to debug.
I treat everything I find in code review as something to integrate into the prompts. Eventually, on a given project, you end up getting correct PRs without manual intervention. That's what they mean. You still have to review your code of course!
No, I think it is normal. If it were easy to gain a mental model of the code simply by reading, then debugging would be trivial. The whole point of debugging is that there are differences between your mental model of the code and what the code is actually doing, that sometimes can't be uncovered unless you step through it line by line even if you're the one who wrote it.
It is why I am a bit puzzled by the people who use an LLM to generate code in anything other than a "tightly scoped" fashion (boilerplate, throwaway code, standalone script, single file, or at the function level). I'm not sure how that makes your job later on any easier if you have even a worse mental model of the code because you didn't even write it. And debugging is almost usually more tedious than writing code, so you've traded off the fun/easy part for a more difficult one. Seems like a faustian deal.
I've found planning to be key here for scaling to arbitrary complex changes.
It's much easier to review larger changes when you've aligned on a Claude generated plan up front.
Skills are also a convenient way for writing self-documenting packages. They solve the problem of teaching the LLM how to use a library.
I have started experimenting with a skills/ directory in my open source software, and then made a plugin marketplace that just pulls them in. It works well, but I don't know how scalable it will be.
https://github.com/juanre/ai-tools
I've been having good results with Roo Code/Cline + Claude Code as the provider. I feel like using CC CLI effectively requires manually recreating Roo/Cline
I really enjoyed reading this. One thought I had on the issue of paths in Claude.md
My concern with hardcoding paths inside a doc, it will likely become outdated as the codebase evolves.
One solution would be to script it and have it run pre commit to regenerate the Claude.md with the new paths.
There probably is potential for even more dev tooling that 1. Ensure reference paths are always correct, 2. Enforces standard for how references are documented in Claude.md (and lints things like length)
Perhaps using some kind of inline documentation standard like jsdoc if it’s a ts file or a naming convention if it’s an Md file
Example:
// @claude.md // For complex … usage or if you encounter a FooBarError, see ${path} for advanced troubleshooting steps
We have a linter that checks for this to help mitigate
You lint the file paths inside Claude.md?
All markdown files, yeah
I feel like these posts are interesting, but become irrelevant quickly. Does anyone actually follow these as guides, or just consume them as feedback for how we wish we could interface with LLMs and the workarounds we currently use?
Right now these are reading like a guide to prolog in the 1980s.
Given that this space is so rapidly evolving, these kinds of posts are helpful just to make sure you aren't missing anything big. I've caught myself doing something the hard way after reading one of these. In this case, the framing is basically man pages for CLIs was a helpful description of sills that gives me some ideas about how to improve interaction with an in-house CLI my co. uses.
Yeah I like to think not everyone can spend their day exploring/tinkering with all these features so it's handy to just snapshot what exists and what works/doesn't.
I wouldn't say I follow them as guides, but I think the field is changing quickly enough that it's good, or at least interesting, to read what's working well for other people.
This one is already out of date. The bit on the top about allocating space in CLAUDE.md for each tool is largely a waste of tokens these days. Use the skills feature.
It's a balance and we use both.
Skills doesn't totally deprecate documenting things in CLAUDE.md but agree that a lot of these can be defined as skills instead.
Skill frontmatter also still sits in the global context so it's not really a token optimization either.
The skill lets you compress the amount loaded to just the briefest description, with the “where do I go to get more info” being implicit. You should use a SKILL.md for evry added tool. At which point, putting instructions in CLAIDE.md becomes redundant and confusing to the LLM.
I wouldn't use as a guide necessarily, but I would use as a way to sync my own findings and see if I have missed something important.
Are "agents" really as useful as the hype makes them out to be?
Most of the time I'm just pasting code blocks directly into raycast and once I've fixed the bug or got the properly transformed code in the shape that I aimed for, then I paste it back into neovim. Next i'm going to try out "opencode"[0], because I've heard some good things about it. For now, I'm happy with my current workflow.
[0] https://github.com/NickvanDyke/opencode.nvim
to be clear that repo is simply a thin neovim plugin for the truly excellent agent TUI opencode
I recommend using it directly instead of via the plugin
Yes, agents are really as useful as the hype makes them out to be. Certainly they’re more useful on bigger code bases but even small ones the agent provides more value than you may expect. I have a small side project probably around 10k loc and the ai agents still plow through tokens. End of the day letting the ai read your entire code base really does let it make more intelligent design decisions. Many other benefits as well. Iteration is much faster without copy paste
I don't understand how people use the `git worktree` workflow. I get that you want to isolate your work, but how do you deal with dev servers, port conflicts and npm installs? When I tried it, it was way more hassle than it was worth.
Yeah it is a mystery to me how folks could also maintain context in more than two sessions. The code review would be brutal.
You’ll also end up dealing with merge conflicts if you haven’t carefully split the work or modularized the code.
I generally like to use it. But I one project in the org which simply can’t work because the internal built system expects a normal .git directory at the root. Means I have to rewrite some of the build code that isn’t aware of this git feature. And yes we use a library to read from git but not the git cli or a more recent compatible one that understands that the current work tree is not the main one.
Agree, depending on the repo and changes it’s hard with local dev servers. It sometimes works well if you don’t need local dockers and want to outsource git workflow to CC as well. Then it can do on that branch whatever it wants and main work is in another worktree with more steering and or docker env.
I have a bash script that creates the worktree, copies env over and changes the ports of containers and the services. I then can proxy the "real" port to any worktree, it's common I'll have 3 worktrees active to switch back and forth
I gave that a try, then I decided to use devcontainers instead, and I find that better, for the reasons you mentioned.
Hooks are underutilized and will be critical for long-running and better agent performance. Excited to release Cupcake in the coming weeks. What I started building when i made the feature request for hooks.
- https://github.com/eqtylab/cupcake
- https://github.com/anthropics/claude-code/issues/712
"It shifts us from a “per-seat” license to “usage-based” pricing, which is a much better model for how we work. It accounts for the massive variance in developer usage (We’ve seen 1:100x differences between engineers)."
This is how we are doing it too, for the same reasons. For now, much easier to administer than trying to figure out who spends enough to get a real Claude Code seat. The other nice thing about using API keys is that you basically never hit rate limits.
Does anyone have any suggestions on making Claude prefer to use project internal abstractions and utility functions? My C++ project has a lot of them. If I just say something like "for I/O and networking code, check IOUtils.h for helpers" then it often doesn't do that. But mentioning all helper functions and classes in the context also seems like a bad idea. What's the best way? Are the new Skills a solution?
Skills seem like the way forward, but Claude still needs to be convinced to activate the skill. If that's not happening reliably, hooks should be able to help.
A sibling comment on hooks mentions some approaches. You could also try leveraging the UserPromptSubmit hook to do some prompt analysis and force relevant skill activation.
I wonder how well a sentence or two in CLAUDE.md, saying to search the local project for examples of similar use cases or use of internal libraries, would work.
Hooks can also be useful for this. If it's using the wrong APIs then can hint on write or block on commit with some lint function that checks for this.
What type of projects are you guys building? I bought Max and these features to try it out to build a more complex project (ROS2) and that does not seem to work out at all… HTML page, yes, embedded project not so much.
With all the LLM coding assistants, you have to get a feel for each model and which extension/interface you're using with them. Not only that, but it's also dependent on your project!
For example, if you're writing a command line tool in Python, it doesn't really matter what model you use since they're all really great at Python (LOL). However, if you're writing a complicated SPA that uses say, Vue 3 with Vite (and some fancy CSS framework) and Python w/FastAPI... You want the "smartest" model that knows about all these things at once (and regularly gets updated knowledge of the latest versions of things). For me, that means Claude Code.
I am cheap though and only pay Anthropic $20/month. This means I run out of Claude Credits every damned week (haha). To work around this problem, I used to use OpenAI's pay-per-use API with gpt5-mini with VS Code's Copilot extension, switching to GPT5-codex (medium) with the Codex extension for more complicated tasks.
Now that I've got more experience, I've figured out that GPT5-codex costs way too much (in API credits) for what you get in nearly all situations. Seriously: Why TF does it use that much "usage". Anyway...
I've tried them all with my very, very complicated collaborative editor (CRDTs), specifically to learn how to better use AI coding assistants. So here's what I do now:
I cannot understate how impressed I am with gpt-oss:120b... It's like 10x faster than gpt5-mini and yet seems to perform just as well. Maybe better, actually because it forces you to narrow your prompts (due to smaller context window). But because it's just so damned fast, that doesn't matter.With Claude Code, it's like magic: You give it a really f'ing complicated thing to troubleshoot or implement and it just goes—and keeps going until it finishes or you run out of tokens! It's a, "the future is now!" experience for sure.
With gpt-oss:120b it's more like having an actual conversation, where the only time you stop typing is when you're reviewing what it did (which you have to do for all the models... Some more than others).
FYI: The worst is Gemini 2.5. I wouldn't even bother! It's such trash, I can't even fathom how Google is trying to pass it off as anything more than a toy. When it decides to actually run (as opposed to responding with, "Failed... Try again"), it'll either hallucinate things that have absolutely nothing to do with your prompt or it'll behave like some petulant middle school kid that pretend to spend a lot of time thinking about something but ultimately does nothing at all.
You do know that you can run Codex with the $20 ChatGPT subscription right? So the token waste doesn't matter that much. It's still slow though.
Tried that. I hit the limit way too fast. Faster than Claude Code, even!
GPT5-codex (medium) is such a token hog for some reason
Just read it and the author say "do not use NEVER" in claude.md then in his sample claude.md there is NEVER the word, am I missing something?
They said don’t __just__ use never, and explain that as meaning you should give an alternative with every “never” clause, which is what their example does :)
Are there any of those CLI clients (coded in plain and simple C, or basic python/perl without 1 billion of expensive dependencies) able to access those 'coding AI' prompt anonymously then rate limited?
If no anonymous access is provided, is there a way to create an account with a noscript/basic (x)html/classic web browsers in order to get an API key secret?
Because I do not use web engines from the "whatng" cartel.
To add insult to injury, my email is self-hosted with IP literals to avoid funding the DNS people which are mostly now in strong partnership with the "whatng" cartel (email with IP literals are "stronger" than SPF since it does the same and more). An email is often required for account registration.
Just my curiosity: Why are you producing so much code? Is it because it is now possible to do so with AI, or because you have a genuine need (solid business usecase) that requires a lot of code?
I just started developing self-hosted services largely with AI.
It wasn't possible before for me to do any of this at this kind of scale. Before, getting stuck on a bug could mean hours, days, or maybe even weeks of debugging. I never made the kind of progress I wanted before.
Many of the things I want, do already exist, but are often older, not as efficient or flexible as they could be, or just plain _look_ dated.
But now I can pump out react/shadcn frontends easily, generate apis, and get going relatively quickly. It's still not pure magic. I'm still hitting issues and such, but they are not these demotivating, project-ending, roadblocks anymore.
I can now move at a speed that matches the ideas I have.
I am giving up something to achieve that, by allowing AI to take control so much, but it's a trade that seems worth it.
Often code in SaaS companies like ours is indeed how we solve customer problems. It's not so much the amount of code but the rate (code per time) we can effectively use to solve problems/build solutions. AI, when tuned correctly, lets us do this faster than ever possible before.
>> Why are you producing so much code?
This is basically a "thinking tax".
If you don't want to think and offload it to llm they burn through a lot of tokens to implement in a non-efficient way something you could often do in 10 lines if you though about it for a few minutes.
I’ve just implemented a proof of concept that involved an API, a MCP server, an Authorization Server, a React frontend, token validation and proof of possession on the client, a CIBA flow for authentication… took a week , and I don’t even know the technologies used very well, it was all TypeScript but I work on JVM languages normally. This was a one off for a customer and I was able to show a fairly complex workflow end to end and what each part involves. I let the LLM write most of it but I understand every line and did have to make manual adjustments (though to be honest, I could easily explain to the LLM what I needed changed and given my experience it would eventually get there.
If you tell me I didn’t really need a LLM to be able to do all that in a week and just some thought and 10 lines of code would do, I suspect you are not really familiar with the latest developments in AI and just vastly underestimates the capabilities they have to do tricky stuff.
>> I don’t even know the technologies used very well
Thats why it took a week with llm. And for you it makes sense as this is new tech.
But if someone knows those technologies - it would still take a week with llm and like 2 days without.
In a large project with decent code structure there can be quite a bit of boilerplate, convention, testing required. Also we are not talking about a 10-line change. More like 10k line feature.
Before LLMs we simply wouldn't implement many of those features since they were not exactly critical and required a lot of time, but now when the required development time is cut signifficantly, they suddenly make sense to implement.
"All my stateless tools (like Jira, AWS, GitHub) have been migrated to simple CLIs." - How do you get Jira on the CLI?
There's an Atlasian cli with Jira support https://developer.atlassian.com/cloud/acli/reference/command...
Cloud only. My employer is still on an ancient data center version. But you can easily write a cli that wraps the REST API.
Jiratui[0] has some support for basic automation. That's probably what OP is using as it is the most poppular Jira cli tool out there.
0: https://github.com/whyisdifficult/jiratui
First search result (on Kagi): https://github.com/ankitpokhrel/jira-cli
Latest version from 2 momths ago, >4700 stars on GitHub
At some point I vibecoded myself everything into cli commands, anything that has API could be a cli command.
Or use this solution and get fine-tuned context generated for each and every task just-in-time : devly.ai
Please stop expecting every engineer on the team to be an ai engineer just to get started with coding agents
Why are so many still using CC and not Codex
If you have no modifications or customization of Claude code then it comes down to a preference for proactivity (codex) or a bit more restraint.
If you are using literally any of Claude Code’s features the experience isn’t close, and regardless of model preference (Claude is my least favorite model by far) you should probably use Claude code. It’s just a much more extensible product for teams.
Which features are preferable to higher quality output?
Losing access to GPT 5 Pro is also a big hit… it is by far the best for reading full files/repos and creating plans (though it also by far has the worst out of the box tooling)
CC has better agent tools and is faster. The ability to switch from plan mode to execution mode and back is huge. Toggling thinking also. And of course they are innovating all of these agentic features like MCP, sub-agents, skills, etc...
Codex writes higher quality code, but is slower and less feature rich. I imagine this will change within months. The jury is still out. Exciting times!
I guess I don’t understand wanting faster and worse for much work, and some of the features like subagents are dubious or like skills and planning mode are minor conveniences over skill files mentioned by agents.md and toggling read only mode or using a plan file. After all those latter features are just conveniences for assembling context.
Maybe CC users haven’t figured out how to parallelize their work because it’s fast enough to just wait or be distracted, and so the Codex waiting seems unbearable.
A lot of the time the code that needs to be written isn't something that requires an extremely powerful model.
I use both at the same time. CC seems to have better access to web and researching capabilities compared to Codex. Maybe I'm not using Codex right or missing something, but it has frequent troubles browsing internet. Also Claude Code is faster. So I use it when I know it can handle the task.
Both. Codex MCP within CC as a second brain. Best of both worlds.
Ecosystem features and cohesion.
What features are preferable to better output quality? (Since you didn’t mention output quality as superior)
right on. i usually just tell it "hey go update this function to do [x]" in horribly misspelled english and then yell at it until it does it right
Sometimes I write with Claude in English and German mixed with really bad typos and it’s amazing how well it works.
When touch typing and talking to someone, I accidentally typed something to claude with my fingers off the home row, e.g. ttoubg kuje tgus ubti tge ckayde cide ternubak. Claude understood it just fine. Didn't even remark on it.
It truly is an idiot savant. It's absurdly good, and then if there is any unaccounted complexity, ugh let's just make the tests stubs.... tests pass now.
Ja find ich auch schön schreckliches Denglisch mit Claude zu reden.
wait until you start conversing with it. It's been a game changer for me how I use Claude CLI. It suits my workflow fine since my sessions are intense in focus I have to bring and I iterate with it; I just haven't found _a way_ where I can give it a large thing to work on and that it will not deviate. I do one focused thing at a time, review, test, alter code and then repeat. With voice mode it's been great since I can talk with it while walking fast on a treadmill. It's bizarre, star trekish, and it works. I wish I could have a stop word with whisper, since I do tend to think long between sentences in this mode, and I wish I could stop it with voice while it's talking, but I found a flow that it doesn't matter that much.
I suggest everyone who can to try the voice mode. https://getvoicemode.com/
Hans@lhoest.eu
[dead]
[dead]
Enshittification needs its Moore's law.
Good article
No thanks. I rather write the code myself that use generated slop. I actually like to code and see little benefit in other peoples copypaste code (thats essentially what ai slop is really)
I feel or have the fear that the world will tumble and crack under the sheer amount of code we produce and can’t be maintained because at one point no one human can understand all the stuff that was written.
At the moment though I also code on and off with an agent. I’m not ready or willing to only vibe code my projects. For one is the fact that I had tons of examples where the agent gaslighted me only to turn around at the last stage. And in some cases the code output was to result focused and didn’t think about the broader general usage. And sure that’s in part because I hold it wrong. Don’t specify 10million markdown files etc. But it’s a feedback loop system. If I don’t trust the results I don’t jump in deeper. And I feel a lot of developers have no issue with jumping ever deeper. Write MCPs now CLIs and describe projects with custom markdown files. But I think we really need both camps. Otherwise we don’t move forward.
> I feel or have the fear that the world will tumble and crack under the sheer amount of code we produce and can’t be maintained because at one point no one human can understand all the stuff that was written.
IMO the best advice in life is try not to be fearful of things that happen to everyone and you can't change.
Good news! What you are afraid of will happen, but it'll happen to everyone all at once, and nothing you can do can change it.
So you no longer need to feel fear. You can skip right on over to resignation. (We have cookies, for we are cooked)