Interesting, I’ve never needed 1M, or even 250k+ context. I’m usually under 100k per request.
About 80% of my code is AI-generated, with a controlled workflow using dev-chat.md and spec.md. I use Flash for code maps and auto-context, and GPT-4.5 or Opus for coding, all via API with a custom tool.
Gemini Pro and Flash have had 1M context for a long time, but even though I use Flash 3 a lot, and it’s awesome, I’ve never needed more than 200k.
For production coding, I use
- a code map strategy on a big repo. Per file: summary, when_to_use, public_types, public_functions. This is done per file and saved until the file changes. With a concurrency of 32, I can usually code-map a huge repo in minutes. (Typically Flash, cheap, fast, and with very good results)
- Then, auto context, but based on code lensing. Meaning auto context takes some globs that narrow the visibility of what the AI can see, and it uses the code map intersection to ask the AI for the proper files to put in context. (Typically Flash, cheap, relatively fast, and very good)
- Then, use a bigger model, GPT 5.4 or Opus 4.6, to do the work. At this point, context is typically between 30k and 80k max.
What I’ve found is that this process is surprisingly effective at getting a high-quality response in one shot. It keeps everything focused on what’s needed for the job.
Higher precision on the input typically leads to higher precision on the output. That’s still true with AI.
For context, 75% of my code is Rust, and the other 25% is TS/CSS for web UI.
Anyway, it’s always interesting to learn about different approaches. I’d love to understand the use case where 1M context is really useful.
Yeah this is the simpler and also effective strategy. A lot of people are building sophisticated AST RAG models. But you really just need to ask Claude to generally build a semantic index for each large-ish piece of code and re-use it when getting context.
You have to make sure the semantic summary takes up significantly less tokens than just reading the code or its just a waste of token/time.
Then have a skill that uses git version logs to perform lazy summary cache when needed.
It seems like a very good use of LLMs. You should write a blog post with detail of your process with examples for people who are not into all AI tools as much. I only use Web UI. Lots of what you are saying is beyond me, but it does sound like clever strategy.
Yeah we all converge to the same workflow, in my ai coding agent I'm working on now, I've added an "index" tool that uses tree-sitter to compress and show the AI a skeleton of a code file.
Fair point, but because I spent a year building and refining my custom tool, this is now the reality for all of my AI requests.
I prompt, press run, and then I get this flow:
dev setup (dev-chat or plan)
code-map (incremental 0s 2m for initial)
auto-context (~20s to 40s)
final AI query (~30s to 2m)
For example, just now, in my Rust code (about 60k LOC), I wanted to change the data model and brainstorm with the AI to find the right design, and here is the auto-context it gave me:
- Reducing 381 context files ( 1.62 MB)
- Now 5 context files ( 27.90 KB)
- Reducing 11 knowledge files ( 30.16 KB)
- Now 3 knowledge files ( 5.62 KB)
The knowledge files are my "rust10x" best practices, and the context files are the source files.
I built myself an AST based solution for that during the last 6 months roughly. I always wondered whether grep and agent-based discovery will be the end of it and thought it just has to be better with a more deterministic approach.
In the end it's hard to measure but personally I feel that my agent rarely misses any context for a given task, so I'm pretty happy with it.
I used a different approach than tree-sitter because I thought I found a nice way to get around having to write language-specific code. I basically use VSCode as a language backend and wrote some logic to basically rebuild the AST tree from VSCode's symbol data and other API.
That allows me to just install the correct language extension and thus enable support for that specific language. The extension has to provide symbol information which most do through LSP.
In the end it was way more effort than just using tree-sitter, however, and I'm thinking of doing a slow migration to that approach sooner or later.
Anyways, I created an extension that spins up an mcp server and provides several tools that basically replace the vanilla discovery tools in my workflow.
The approach is similar to yours, I have an overview tool which runs different centrality ranking metrics over the whole codebase to get the most important symbols and presents that as an architectural overview to the LLM.
Then I have a "get-symbol-context" tool which allows the AI to get all the information that the AST holds about a single symbol, including a parameter to include source code which completely replaces grepping and file reading for me.
The tool also specifies which other symbols call the one in question and which others it calls, respectively.
But yeah, sorry for this being already a quite long comment, if you want to give it a try, I published it on the VSCode marketplace a couple of days ago, and it's basically free right now, although I have to admit that I still want to try to earn a little bit of money with it at some point.
Right now, the daily usage limit is 2000 tool calls per day, which should be enough for anybody.
I looked at your solution and extension README, and it's very interesting and well thought out.
The fact that you've been using it for six months and that it performs well says a lot. At the end of the day, that's what counts.
I like your idea of piggybacking on top of the LSP services, and I can imagine that this was quite a bit of work. Doing it as an MCP server makes it usable across different tools.
I also really like the name "Context Master."
In my case, it's much more niche since it's for the tool I built. Though it's open source, the key difference is that the "indexing" is only agentic at this point.
I can see value in mixing the two. LSP integration scares me because of the amount of work involved, and tree-sitter seems like a good path.
In that case, in the code map, for each item, there could be both the LLM response info and some deterministic info, for example, from tree-sitter.
That being said, the current approach works so well that I think I am going to keep using and fine-tuning it for a while, and bring in deterministic context only when or if I need it.
Anyway, what you built looks great. If it works, that's great.
This is really interesting; ive done very high level code maps but the entire project seems wild, it works?
So, small model figures out which files to use based on the code map, and then enriches with snippets, so big model ideally gets preloaded with relevant context / snippets up front?
So, I have a pro@coder/.cache/code-map/context-code-map.json.
I also have a `.tmpl-code-map.jsonl` in the same folder so all of my tasks can add to it, and then it gets merged into context-code-map.json.
I keep mtime, but I also compute a blake3 hash, so if mtime does not match, but it is just a "git restore," I do not redo the code map for that file. So it is very incremental.
Then the trick is, when sending the code map to AI, I serialize it in a nice, simple markdown format.
- path/to/file.rs
- summary: ...
- when to use: ...
- public types: .., .., ..
- public functions: .., .., ..
- ...
So the AI does not have to interpret JSON, just clean, structured markdown.
Funny, I worked on this addition to my tool for a week, planning everything, but even today, I am surprised by how well it works.
I have zero sed/grep in my workflow. Just this.
My prompt is pro@coder/coder-prompt.md, the first part is YAML for the globs, and the second part is my prompt.
There is a TUI, but all input and output are files, and the TUI is just there to run it and see the status.
> - a code map strategy on a big repo. Per file: summary, when_to_use, public_types, public_functions. This is done per file and saved until the file changes. With a concurrency of 32, I can usually code-map a huge repo in minutes. (Typically Flash, cheap, fast, and with very good results)
Thanks, but why use any AI to generate this? I would say: you document your functions-in-code, types are provided from the compiler service, so it should all be deterministically available in seconds iso minutes, without burning tokens. Am I missing something?
- Using a tree-sitter/AST-like approach, I could extract types, functions, and perhaps comments, and put them into an index map.
- Cons:
- The tricky part of this approach is that what I extract can be pretty large per file, for example, comments.
- Then, I would probably need an agentic synthesis step for those comments anyway.
2) Agentic
- Since Flash is dirt cheap, I wanted to experiment and skip #1, and go directly to #2.
- Because my tool is built for concurrency, when set to 32, it's super fast.
- The price is relatively low, perhaps $1 or $2 for 50k LOC, and 60 to 90 seconds, about 30 to 45 minutes of AI work.
- What I get back is relatively consistent by file, size-wise, and it's just one trip per file.
So, this is why I started with #2.
And then, the results in real coding scenarios have been astonishing.
Way above what I expected.
The way those indexes get combined with the user prompt gets the right files 95% of the time, and with surprisingly high quality.
So, I might add deterministic aspects to it, but since I think I will need the agentic step anyway, I have deprioritized it.
I think you've kind of hit on the more successful point here, which is that you should be keeping things focused in a sufficiently focused area to have better success and not necessarily needing more context.
Well, out of all the workflows I have seen, this one is rather nice, might give it a try.
I imagine if the context were being commited and kept up-to-date with CI would work for others to use as well.
However, I'm a little confused on the autocontext/globs narrowing part. Do you, the developer, provide them? Or you feed the full code map to flash + your prompt so it returns the globs based on your prompt?
Also, in general, is your map of a file relatively smaller than the file itself, even for very small files?
- The ..-code-map.json files are per "developer folder," which would create too many conflicts if they were kept in Git.
- I have two main globs, which are lists of globs: knowledge_globs and context_globs. Knowledge can be absolute and should be relatively static. context_globs have to be relative to the workspace, since they are the working files.
- As a dev, you provide them in the top YAML section of the coder-prompt.md.
- The auto-context sub-agent calls the code-map sub-agent. Sub-agents can add to or narrow the given globs, and that is the goal of the auto-context agent.
It looks complicated, but it actually works like a charm.
Hopefully, I answered some of your questions.
I need to make a video about it.
But regardless, I really think it's not about the tools, it's about the techniques. This is where the true value is.
This is interesting but don't you worry that you're competing with entire companies (e.g. Anthropic) and thus it's a losing battle? Since you're re-implementing a bunch of stuff they either do in their harness or have decided it was better not to do?
I think it's worth remembering that for any offering like that, it necessarily needs to be ~one-size-fits-all, while what you come up with.. doesn't.
They're solving a different problem than you. So I think it's very plausible that you could come up with something that, for your use case, performs considerably better than their "defaults".
Personally I don't see aipack's pro@coder and other approaches (claude code, cursor, copilot, etc...) as competitors anymore. I use both approaches to solve different problems. I keep using the agentic solutions (claude code style) for more operational tasks, a bit like "smart interfaces to terminal", and pro@coder for coding / engineering tasks where I need a much tighter control over long running work sessions.
Your code map compresses signal on the context side. Same principle applies on the prompt side: prompts that front-load specifics (file, error, expected behavior) resolve in 1-2 turns. Vague ones spiral into 5-6. 1M context doesn't change that — it just gives you more room for the spiral.
This is fascinating. I feel like this is converging into the concept of a traditional "IDE". So much of your setup reminds me of IDEs indexing, doing static analysis, building ASTs, etc. before a developer starts writing code.
> Standard pricing now applies across the full 1M window for both models, with no long-context premium. Media limits expand to 600 images or PDF pages.
For Claude Code users this is huge - assuming coherence remains strong past 200k tok.
Is it ever useful to have a context window that full? I try to keep usage under 40%, or about 80k tokens, to avoid what Dex Horthy calls the dumb zone in his research-plan-implement approach. Works well for me so far.
I'd been on Codex for a while and with Codex 5.2 I:
1) No longer found the dumb zone
2) No longer feared compaction
Switching to Opus for stupid political reasons, I still have not had the dumb zone - but I'm back to disliking compaction events and so the smaller context window it has, has really hurt.
I hope they copy OpenAI's compaction magic soon, but I am also very excited to try the longer context window.
● RESEARCH. Don't code yet. Let the agent scan the files first. Docs lie. Code doesn't.
● PLAN. The agent writes a detailed step-by-step plan. You review and approve the plan, not just the output. Dex calls this avoiding "outsourcing your thinking." The plan is where intent gets compressed before execution starts.
● IMPLEMENT. Execute in a fresh context window.
The meta-principle he calls Frequent Intentional Compaction: don't let the chat run long. Ask the agent to summarize state, open a new chat with that summary, keep the model in the smart zone.
For me, it's less about being able to look back -800k tokens. It's about being able to flow a conversation for a lot longer without forcing compaction. Generally, I really only need the most recent ~50k tokens, but having the old context sitting around is helpful.
When running long autonomous tasks it is quite frequent to fill the context, even several times. You are out of the loop so it just happens if Claude goes a bit in circles, or it needs to iterate over CI reds, or the task was too complex. I'm hoping a long context > small context + 2 compacts.
It's kind of like having a 16 gallon gas tank in your car versus a 4 gallon tank. You don't need the bigger one the majority of the time, but the range anxiety that comes with the smaller one and annoyance when you DO need it is very real.
Since I'm yet to seriously dive into vibe coding or AI-assisted coding, does the IDE experience offer tracking a tally of the context size? (So you know when you're getting close or entering the "dumb zone")?
I never use these giant context windows. It is pointless. Agents are great at super focused work that is easy to re-do. Not sure what is the use case for giant context windows.
Maxing out context is only useful if all the information is directly relevant and tightly scoped to the task. The model's performance tends to degrade with too much loosely related data, leading to more hallucinations and slower results. Targeted chunking and making sure context stays focused almost always yields better outcomes unless you're attempting something atypical, like analyzing an entire monorepo in one shot.
I've used it many times for long-running investigations. When I'm deep in the weeds with a ton of disassembly listings and memory dumps and such, I don't really want to interrupt all of that with a compaction or handoff cycle and risk losing important info. It seems to remain very capable with large contexts at least in that scenario.
After running a context window up high, probably near 70% on opus 4.6 High and watching it take 20% bites out of my 5hr quota per prompt I've been experimenting with dumping context after completing a task. Seems to be working ok. I wonder if I was running into the long context premium. Would that apply to Pro subs or is just relevant to api pricing?
I haven't hit the "dumb zone" any more since two months. I think this talk is outdated.
I'm using CC (Opus) thinking and Codex with xhigh on always.
And the models have gotten really good when you let them do stuff where goals are verifiable by the model. I had Codex fix a Rust B-rep CSG classification pipeline successfully over the course of a week, unsupervised. It had a custom STEP viewer that would take screenshots and feed them back into the model so it could verify the progress resp. the triangle soup (non progress) itself.
Codex did all the planning and verification, CC wrote the code.
This would have not been possible six months ago at all from my experience.
Maybe with a lot of handholding; but I doubt it (I tried).
I mean both the problem for starters (requires a lot of spatial reasoning and connected math) and the autonomous implementation. Context compression was never an issue in the entire session, for either model.
I mean, try using copilot on any substantial back-end codebase and watch it eat 90+% just building a plan/checklist. Of course copilot is constrained to 120k I believe? So having 10x that will blow open up some doors that have been closed for me in my work so far.
That said, 120k is pleeenty if you’re just building front-end components and have your API spec on hand already.
Fwiw put a copy of the game folder in a directory and tell claude to extract game files and dissasemble the game in preparation for questions about the game.
Context decay is noticeable within 3 messages, nearly every time. Maybe not substantial, but definitely noticeable.
It’s lead to me starting new chats with bigger and bigger starting ‘summary, prompts to catch the model up while refreshing it. Surely there’s a way to automate that technique.
Adding web search doesn't necessarily lead to better information at any context.
In my experience the model will assume the web results are the answer even if the search engine returns irrelevant garbage.
For example you ask it a question about New Jersey law and the web results are about New York or about "many states" it'll assume the New York info or "many states" info is about New Jersey.
I think ChatGPT has a huge advantage here. They have been collecting realistic multi-turn conversational data at a much larger scale. And generally their models appear to be more coherent with larger contexts for general purpose stuff.
Agreed, there's no getting around the "break it into smaller contexts" problem that lies between us and generally useful AI.
It'll remain a human job for quite a while too. Separability is not a property of vector spaces, so modern AIs are not going to be good at it. Maybe we can manage something similar with simplical complexes instead. Ideally you'd consult the large model once and say:
> show me the small contexts to use here, give me prompts re: their interfaces with their neighbors, and show me which distillations are best suited to those tasks
...and then a network of local models could handle it from there. But the providers have no incentive to go in that direction, so progress will likely be slow.
That’s not context decay, that’s training data ambiguity. So much misinformation, nerfs, buffs, changes that an LLM can not keep up given the training time required. Do it for a game that has been stable and it knows its stuff.
What were you asking about PoE 2? So far my _general_ experience with asking LLMs about ARPGs has been meh. Except for Diablo 2 but I think that’s just because Diablo 2 has been heavily discussed for ~25 years.
I've been using the 1M window at work through our enterprise plan as I'm beginning to adopt AI in my development workflow (via Cline). It seems to have been holding up pretty well until about 700k+. Sometimes it would continue to do okay past that, sometimes it started getting a bit dumb around there.
(Note that I'm using it in more of a hands-on pair-programming mode, and not in a fully-automated vibecoding mode.)
The quality with the 1M window has been very poor for me, specifically for coding tasks. It constantly forgets stuff that has happened in the existing conversation. n=1, ymmv
Yes, especially with shifts in focus of a long conversation. But given the high error rates of Opus 4.6 the last few weeks it is possibly due to other factors. Conversational and code prodding has been essential.
Well, the question is what is contributing to the usage. Because as the context grows, the amount of input tokens are increasing. A model call with 800K token as input is 8 times more expensive than a model call with 100K tokens as input. Especially if we resume a conversation and caching does not hit, it would be very expensive with API pricing.
It’s interesting because my career went from doing higher level language (Python) to lower language (C++ and C). Opus and the like is amazing at Python, honestly sometimes better than me but it does do some really stupid architectural decisions occasionally. But when it comes to embedded stuff, it’s still like a junior engineer. Unsure if that will ever change but I wonder if it’s just the quality and availability of training data. This is why I find it hard to believe LLMs will replace hardware engineers anytime soon (I was a MechE for a decade).
As someone who did Python professionally from a software engineering perspective, I've actually found Python to be pretty crappy really: unaware of _good_ idioms living outside tutorials and likely 90% of Python code out there that was simply hacked together quickly.
I have not tested, but I would expect more niche ecosystems like Rust or Haskell or Erlang to have better overall training set (developer who care about good engineering focus on them), and potentially produce the best output.
For C and C++, I'd expect similar situation with Python: while not as approachable, it is also being pushed on beginning software engineers, and the training data would naturally have plenty of bad code.
I've found it's ok at Rust. I think a lot of existing Rust code is high quality and also the stricter Rust compiler enforces that the output of the LLM is somewhat reasonable.
I think the combinatorial space is just too much. When I did web dev it was mostly transforming HTML/JSON from well-defined type A to well-defined type B. Everything is in text. There's nothing to reason about besides what is in the prompt itself. But constructing and maintaining a mental model of a chip and all of its instructions and all of the empirical data from profiling is just too much for SOTA to handle reliably.
Unfortunately, I’ve found it’s really good at Wayland and OpenGL. It even knows how to use Clutter and Meta frameworks from the Gnome Mutter stack.
Makes me wonder why I learned this all in the first place.
Is there a writeup anywhere on what this means for effective context? I think that many of us have found that even when the context window was 100k tokens the actual usable window was smaller than that. As you got closer to 100k performance degraded substantially. I'm assuming that is still true but what does the curve look like?
> As you got closer to 100k performance degraded substantially
In practice, I haven't found this to be the case at all with Claude Code using Opus 4.6. So maybe it's another one of those things that used to be true, and now we all expect it to be true.
And of course when we expect something, we'll find it, so any mistakes at 150k context use get attributed to the context, while the same mistake at 50k gets attributed to the model.
My personal experience is that Opus 4.6 degrades after a while but the degradation is more subtle and less catastrophic than in the past. I still aggressively clear sessions to keep it sharp though.
Personally, even though performance up to 200k has improved a lot with 4.5 and 4.6, I still try to avoid getting up there — like I said in another comment, when I see context getting up to even 100k, I start making sure I have enough written to disk to type /new, pipe it the diff so far, and just say “keep going.” I feel like the dropoff starts around maybe 150k, but I could be completely wrong. I thought it was funny that the graph in the post starts at 256k, which convenient avoids showing the dropoff I'm talking about (if it's real).
I mentioned this at work but context still rots at the same rate. 90k tokens consumed has just as bad results in 100k context window or 1M.
Personally, I’m on a 6M+ line codebase and had no problems with the old window. I’m not sending it blindly into the codebase though like I do for small projects. Good prompts are necessary at scale.
Isn't transformer attention quadratic in complexity in terms of context size? In order to achieve 1M token context I think these models have to be employing a lot of shortcuts.
I'm not an expert but maybe this explains context rot.
Nope, there’s no tricks unless there’s been major architectural shifts I missed. The rot doesn’t come from inference tricks to try to bring down quadratic complexity of the KV cache. Task performance problems are generally a training problem - the longer and larger the data set, the fewer examples you have to train on it. So how do you train the model to behave well - that’s where the tricks are. I believe most of it relies on synthetically generated data if I’m not mistaken, which explains the rot.
Those sorts of volume discounts are what you do when you're trying to incentivize more consumption. Anthropic already has more demand then they're logistically able to serve, at the moment (look at their uptime chart, it's barely even 1 9 of reliability). For them, 1 user consuming 5 units of compute is less attractive than 5 users consuming 1 unit.
They would probably implement _diminishing_-value pricing if pure pricing efficiency was their only concern.
5x Max is the plan I use because the Pro plan limits out so quickly. I don't use Claude full-time, but I do need Claude Code, and I do prefer to use Opus for everything because it's focused and less chatty.
Claude Code 2.1.75 now no longer delineates between base Opus and 1M Opus: it's the same model. Oddly, I have Pro where the change supposedly only for Max+ but am still seeing this to be case.
EDIT: Don't think Pro has access to it, a typical prompt just hit the context limit.
The removal of extra pricing beyond 200k tokens may be Anthropic's salvo in the agent wars against GPT 5.4's 1M window and extra pricing for that.
Opus 4.6 is nuts. Everything I throw at it works. Frontend, backend, algorithms—it does not matter.
I start with a PRD, ask for a step-by-step plan, and just execute on each step at a time. Sometimes ideas are dumb, but checking and guiding step by step helps it ship working things in hours.
It was also the first AI I felt, "Damn, this thing is smarter than me."
The other crazy thing is that with today's tech, these things can be made to work at 1k tokens/sec with multiple agents working at the same time, each at that speed.
I wish I had this kind of experience. I threw a tedious but straightforward task at Claude Code using Opus 4.6 late last week: find the places in a React code base where we were using useState and useEffect to calculate a value that was purely dependent on the inputs to useEffect, and replace them with useMemo. I told it to be careful to only replace cases where the change did not introduce any behavior changes, and I put it in plan mode first.
It gave me an impressive plan of attack, including a reasonable way to determine which code it could safely modify. I told it to start with just a few files and let me review; its changes looked good. So I told it to proceed with the rest of the code.
It made hundreds of changes, as expected (big code base). And most of them were correct! Except the places where it decided to do things like put its "const x = useMemo(...)" call after some piece of code that used the value of "x", meaning I now had a bunch of undefined variable references. There were some other missteps too.
I tried to convince it to fix the places where it had messed up, but it quickly started wanting to make larger structural changes (extracting code into helper functions, etc.) rather than just moving the offending code a few lines higher in the source file. Eventually I gave up trying to steer it and, with the help of another dev on my team, fixed up all the broken code by hand.
It probably still saved time compared to making all the changes myself. But it was way more frustrating.
One tip I have is that once you have the diff you want to fix, start a new session and have it work on the diff fresh. They’ve improved this, but it’s still the case that the farther you get into context window, the dumber and less focused the model gets. I learned this from the Claude Code team themselves, who have long advised starting over rather than trying to steer a conversation that has started down a wrong path.
I have heard from people who regularly push a session through multiple compactions. I don’t think this is a good idea. I virtually never do this — when I see context getting up to even 100k, I start making sure I have enough written to disk to type /new, pipe it the diff so far, and just say “keep going.” I learned recently that even essentials like the CLAUDE.md part of the prompt get diluted through compactions. You can write a hook to re-insert it but it's not done by default.
This fresh context thing is a big reason subagents might work where a single agent fails. It’s not just about parallelism: each subagent starts with a fresh context, and the parent agent only sees the result of whatever the subagent does — its own context also remains clean.
Same here. I don't understand how people leave it running on an "autopilot" for long periods of time. I still use it interactively as an assistant, going back and forth and stepping in when it makes mistakes or questionable architectural decisions. Maybe that workflow makes more sense if you're not a developer and don't have a good way to judge code quality in the first place.
There's probably a parallel with the CMSes and frameworks of the 2000s (e.g. WordPress or Ruby on Rails). They massively improved productivity, but as a junior developer you could get pretty stuck if something broke or you needed to implement an unconventional feature. I guess it must feel a bit similar for non-developers using tools like Claude Code today.
Branch first so you can just undo. I think this would have worked with sub agents and /loop maybe? Write all items to change to a todo.md. Have it split up the work with haiku sub agents doing 5-10 changes at a time, marking the todos done, and /loop until all are done. You’ll succeed I suspect. If the main claude instance compacts its context - stop and start from where you left off.
What kinds of things are you building? This is not my experience at all.
Just today I asked Claude using opus 4.6 to build out a test harness for a new dynamic database diff tool. Everything seemed to be fine but it built a test suite for an existing diff tool. It set everything up in the new directory, but it was actually testing code and logic from a preexisting directory despite the plan being correct before I told it to execute.
I started over and wrote out a few skeleton functions myself then asked it write tests for those to test for some new functionality. Then my plan was to the ask it to add that functionality using the tests as guardrails.
Well the tests didn’t actually call any of the functions under test. They just directly implemented the logic I asked for in the tests.
After $50 and 2 hours I finally got something working only to realize that instead of creating a new pg database to test against, it found a dev database I had lying around and started adding tables to it.
When I managed to fix that, it decided that it needed to rebuild multiple docker components before each test and test them down after each one.
After about 4 hours and $75, I managed to get something working that was probably more code than I would have written in 4 hours, but I think it was probably worse than what I would have come up with on my own. And I really have no idea if it works because the day was over and I didn’t have the energy left to review it all.
We’ve recently been tasked at work with spending more money on Claude (not being more productive the metric is literally spending more money) and everyone is struggling to do anything like what the posts on HN say they are doing. So far no one in my org in a very large tech company has managed to do anything very impressive with Claude other than bringing down prod 2 days ago.
Yes I’m using planning mode and clearing context and being specific with requirements and starting new sessions, and every other piece of advice I’ve read.
I’ve had much more luck using opus 4.6 in vs studio to make more targeted changes, explain things, debug etc… Claude seems too hard to wrangle and it isn’t good enough for you to be operating that far removed from the code.
Similar experience. I use these AI tools on a daily basis. I have tons of examples like yours. In one recent instance I explicitly told it in the prompt to not use memcpy, and it used memcpy anyway, and generated a 30-line diff after thinking for 20 minutes. In that amount of time I created a 10-line diff that didn't use memcpy.
I think it's the big investors' extremely powerful incentives manifesting in the form of internet comments. The pace of improvement peaked at GPT-4. There is value in autocomplete-as-a-service, and the "harnesses" like Codex take it a lot farther. But the people who are blown away by these new releases either don't spend a lot of time writing code, or are being paid to be blown away. This is not a hockey stick curve. It's a log curve.
Bigger context windows are a welcome addition. And stuff like JSON inputs is nice too. But these things aren't gonna like, take your SWE job, if you're any good. It's just like, a nice substitute for the Google -> Stack Overflow -> Copy/Paste workflow.
Curious what language and stack. And have people at your company had marginally more success with greenfield projects like prototypes? I guess that’s what you’re describing, though it sounds like it’s a directory in a monorepo maybe?
You probably just don't have the hang of it yet. It's very good but it's not a mind reader and if you have something specific you want, it's best to just articulate that exactly as best you can ("I want a test harness for <specific_tool>, which you can find <here>"). You need to explain that you want tests that assert on observable outcomes and state, not internal structure, use real objects not mocks, property based testing for invariants, etc. It's a feedback loop between yourself and the agent that you must develop a bit before you start seeing "magic" results. A typical session for me looks like:
- I ask for something highly general and claude explores a bit and responds.
- We go back and forth a bit on precisely what I'm asking for. Maybe I correct it a few times and maybe it has a few ideas I didn't know about/think of.
- It writes some kind of plan to a markdown file. In a fresh session I tell a new instance to execute the plan.
- After it's done, I skim the broad strokes of the code and point out any code/architectural smells.
- I ask it to review it's own work and then critique that review, etc. We write tests.
Perhaps that sounds like a lot but typically this process takes around 30-45 minutes of intermittent focus and the result will be several thousand lines of pretty good, working code.
I find that Opus misses a lot of details in the code base when I want it to design a feature or something. It jumps to a basic solution which is actually good but might affect something elsewhere.
GPT 5.4 on codex cli has been much more reliable for me lately. I used to have opus write and codex review, I now to the opposite (I actually have codex write and both review in parallel).
So on the latest models for my use case gpt > opus but these change all the time.
Edit: also the harness is shit. Claude code has been slow, weird and a resource hog. Refuses to read now standardized .agents dirs so I need symlink gymnastics. Hides as much info as it can… Codex cli is working much better lately.
Codex CLI is so much more pleasant to use than CC. I cancelled my CC subscription after the OpenCode thing, but somewhat ironically have recently found myself naturally trying the native Codex CLI client first more often over OpenCode.
Kinda funny how you don't actually need to use coercion if you put in the engineering work to build a product that's competitive on its own technical merits...
My experience is that it gets you 80-90% of the way at 20x the speed, but coaxing it into fixing the remaining 10-20% happens at a staggeringly slow speed.
All programming is like this to some extent, but Claude's 80/20 behavior is so much more extreme. It can almost build anything in 15-30 minutes, but after those 15-30 minutes are up, it's only "almost built". Then you need to spend hours, days, maybe even weeks getting past the "almost".
Big part of why everyone seems to be vibe coding apps, but almost nobody seems to be shipping anything.
I had Opus 4.6 running on a backend bug for hours. It got nowhere. Turned out the problem was in AWS X-ray swizzling the fetch method and not handling the same argument types as the original, which led to cryptic errors.
I had Opus 4.6 tell me I was "seeing things wrong" when I tried to have it correct some graphical issues. It got stuck in a loop of re-introducing the same bug every hour or so in an attempt to fix the issue.
I'm not disagreeing with your experience, but in my experience it is largely the same as what I had with Opus 4.5 / Codex / etc.
Haha, reminds me of an unbelievably aggravating exchange with Codex (GPT 5.4 / High) where it was unflinchingly gaslighting me about undesired behavior still occurring after a change it made that it was adamant simply could not be happening.
It started by insisting I was repeatedly making a typo and still would not budge even after I started copy/pasting the full terminal history of what I was entering and the unabridged output, and eventually pivoted to darkly insinuating I was tampering with my shell environment as if I was trying to mislead it or something.
Ultimately it turned out that it forgot it was supposed to be applying the fixes to the actual server instead of the local dev environment, and had earlier in the conversation switched from editing directly over SSH to pushing/pulling the local repo to the remote due to diffs getting mangled.
I am starting to believe it’s not OPUS but developers getting better at using LLMs across the board. And not realizing they are just getting much better at using these tools.
I also thought it was OPUS 4.5 (also tested a lot with 4.6) and then in February switched to only using auto mode in the coding IDEs. They do not use OPUS (most of the times), and I’m ending up with a similar result after a very rough learning curve.
Now switching back to OPUS I notice that I get more out of it, but it’s no longer a huge difference. In a lot of cases OPUS is actually in the way after learning to prompt more effectively with cheaper models.
The big difference now is that I’m just paying 60-90$ month for 40-50hrs of weekly usage… while I was inching towards 1000$ with OPUS. I chose these auto modes because they don’t dig into usage based pricing or throttling which is a pretty sweet deal.
Seems commonly used in Big Tech - first time I heard it was in my current job. Now it's seared into my brain since it's used so much. Among many other acronyms which I won't bore you with.
I had been able to get it into the classic AI loop once.
It was about a problem with calculation around filling a topographical water basin with sedimentation where calculation is discrete (e.g. turn based) and that edge case where both water and sediments would overflow the basin; To make the matter simple, fact was A, B, C, and it oscillated between explanation 1 which refuted C, explanation 2 which refuted A and explanation 3 that refuted B.
I'll give it to opus training stability that my 3 tries using it all consistently got into this loop, so I decided to directly order it to do a brute force solution that avoided (but didn't solve) this problem.
I did feel like with a human, there's no way that those 3 loop would happen by the second time. Or at least the majority of us. But there is just no way to get through to opus 4.6
Opus helped me brick my RPi CM4 today. It glibly apologized for telling to use an e instead of a 6 in a boot loader sequence.
Spent an hour or so unraveling the mess. My feeling are growing more and more conflicted about these tools. They are here to stay obviously.
I’m honestly uncertain about the junior engineers I’m working with who are more productive than they might be otherwise, but are gaining zero (or very little) experience. It’s like the future is a world where the entire programming sphere is dominated by the clueless non technical management that we’ve all had to deal with in small proportion a time or two.
I've seen a few instances of where Claude showed me a better way to do something and many many more instances of where it fails miserably.
Super simple problem :
I had a ZMK keyboard layout definition I wanted it to convert it to QMK for a different keyboard that had one key less so it just had to trim one outer key. It took like 45 minutes of back and forth to get it right - I could have done it in 30 min manually tops with looking up docs for everything.
Capability isn't the impressive part it's the tenacity/endurance.
Codex doesn't seem that far behind. I use the top model available for api key use and its gotten faster this month even on the max effort level (not like a cheetah - more like not so damn painful anymore). Plus, it also forks agents in parallel - for speed & to avoid polluting the main context. I.e. it will fork explorer agents while investigating (kind of amusing because they're named after famous scientists).
We're never getting UBI. See the latest interview with the Palantir CEO where he talks about white collar workers having to take more hands-on jobs that they may not feel as satisfied with. IE - tending their manors and compounds.
RIP widespread middle class. It was a good 80-year run.
Just yesterday I asked it to repeat a very simple task 10 times. It ended up doing it 15 times. It wasn't a problem per se, just a bit jarring that it was unable to follow such simple instructions (it even repeated my desire for 10 repetitions at the start!).
I’ll put out a suggestion you pair with codex or deepthink for audit and review - opus is still prone to … enthusiastic architectural decisions. I promise you will be at least thankful and at most like ‘wtf?’ at some audit outputs.
Also shout out to beads - I highly recommend you pair it with beads from yegge: opus can lay out a large project with beads, and keep track of what to do next and churn through the list beautifully with a little help.
Opus 4.6 is AGI in my book. They won’t admit it, but it’s absolutely true. It shows initiative in not only getting things right but also adding improvements that the original prompt didn't request that match the goals of the job.
Not even close. There are still tons of architectural design issues that I'd find it completely useless at, tons of subtle issues it won't notice.
I never run agents by themselves; every single edit they do is approved by me. And, I've lost track of the innumerable times I've had to step in and redirect them (including Opus) to an objectively better approach. I probably should keep a log of all that, for the sake of posterity.
I'll grant you that for basic implementation of a detailed and well-specced design, it is capable.
I don’t know if Opus is AGI but on a broader note, that’s how we will get AGI. Not some consciousness like people are expecting. It’s just going to be chatbot that’s very hard to stump and starts making actual scientific breakthroughs and solving long standing problems.
I'm late to the party and I'm just getting started with Antrophic models. I have been finding Sonnet decent enough, but it seems to have trouble naming variables correctly (it's not just that most names are poor and undescriptive, sometimes it names it wrong, confusing) or sometimes unnecessarily declaring, re-declaring variables, encoding, decoding, rather than using the value that's already there etc. Is Opus better at this?
The replies to this really make me think that some people are getting left behind the AI age. Colleges are likely already teaching how to prompt, but a lot of existing software devs just don't get it. I encourage people who aren't having success with AI to watch some youtube videos on best practices.
I'm working in a codebase of 200+ "microservices", separate repos, each deployed as multiple FaaS, CQRS-style. None of it my choice, everything precedes me, many repos I know nothing of.
Little to no code re-use between them.
Any trace of "business logic" is so distributed in multiple repos, that I have no possible use of LLM codegen, unless I can somehow feed it ALL the codebase.
I've tried generating some tests, but they always miss the mark, as in the system under test is almost always the wrong one.
I guess LLM are cool for greenfield, but when the brownfield is really brown, there's no use for LLMs.
This is super exciting. I've been poking at it today, and it definitely changes my workflow -- I feel like a full three or four hour parallel coding session with subagents is now generally fitting into a single master session.
The stats claim Opus at 1M is about like 5.4 at 256k -- these needle long context tests don't always go with quality reasoning ability sadly -- but this is still a significant improvement, and I haven't seen dramatic falloff in my tests, unlike q4 '25 models.
p.s. what's up with sonnet 4.5 getting comparatively better as context got longer?
Did it get better? I used sonnet 4.5 1m frequently and my impression was that it was around the same performance but a hell of a lot faster since the 1m model was willing to spends more tokens at each step vs preferring more token-cautious tool calls.
Readit News
About 80% of my code is AI-generated, with a controlled workflow using dev-chat.md and spec.md. I use Flash for code maps and auto-context, and GPT-4.5 or Opus for coding, all via API with a custom tool.
Gemini Pro and Flash have had 1M context for a long time, but even though I use Flash 3 a lot, and it’s awesome, I’ve never needed more than 200k.
For production coding, I use
- a code map strategy on a big repo. Per file: summary, when_to_use, public_types, public_functions. This is done per file and saved until the file changes. With a concurrency of 32, I can usually code-map a huge repo in minutes. (Typically Flash, cheap, fast, and with very good results)
- Then, auto context, but based on code lensing. Meaning auto context takes some globs that narrow the visibility of what the AI can see, and it uses the code map intersection to ask the AI for the proper files to put in context. (Typically Flash, cheap, relatively fast, and very good)
- Then, use a bigger model, GPT 5.4 or Opus 4.6, to do the work. At this point, context is typically between 30k and 80k max.
What I’ve found is that this process is surprisingly effective at getting a high-quality response in one shot. It keeps everything focused on what’s needed for the job.
Higher precision on the input typically leads to higher precision on the output. That’s still true with AI.
For context, 75% of my code is Rust, and the other 25% is TS/CSS for web UI.
Anyway, it’s always interesting to learn about different approaches. I’d love to understand the use case where 1M context is really useful.
You have to make sure the semantic summary takes up significantly less tokens than just reading the code or its just a waste of token/time.
Then have a skill that uses git version logs to perform lazy summary cache when needed.
Deleted Comment
Here's the implementation for the interested: https://github.com/tontinton/maki/blob/main/maki-code-index%...
I've always wanted to explore how to fit tree-sitter into this workflow. It's great to know that this works well too.
Thanks for sharing the code.
(Here is the AIPack runtime I built, MIT: https://github.com/aipack-ai/aipack), and here is the code for pro@coder (https://github.com/aipack-ai/packs-pro/tree/main/pro/coder) (AIPack is in Rust, and AI Packs are in md / lua)
i said well yeah, but its too sophiscated to be practical
I prompt, press run, and then I get this flow: dev setup (dev-chat or plan) code-map (incremental 0s 2m for initial) auto-context (~20s to 40s) final AI query (~30s to 2m)
For example, just now, in my Rust code (about 60k LOC), I wanted to change the data model and brainstorm with the AI to find the right design, and here is the auto-context it gave me:
- Reducing 381 context files ( 1.62 MB)
- Now 5 context files ( 27.90 KB)
- Reducing 11 knowledge files ( 30.16 KB)
- Now 3 knowledge files ( 5.62 KB)
The knowledge files are my "rust10x" best practices, and the context files are the source files.
(edited to fix formatting)
In the end it's hard to measure but personally I feel that my agent rarely misses any context for a given task, so I'm pretty happy with it.
I used a different approach than tree-sitter because I thought I found a nice way to get around having to write language-specific code. I basically use VSCode as a language backend and wrote some logic to basically rebuild the AST tree from VSCode's symbol data and other API.
That allows me to just install the correct language extension and thus enable support for that specific language. The extension has to provide symbol information which most do through LSP.
In the end it was way more effort than just using tree-sitter, however, and I'm thinking of doing a slow migration to that approach sooner or later.
Anyways, I created an extension that spins up an mcp server and provides several tools that basically replace the vanilla discovery tools in my workflow.
The approach is similar to yours, I have an overview tool which runs different centrality ranking metrics over the whole codebase to get the most important symbols and presents that as an architectural overview to the LLM.
Then I have a "get-symbol-context" tool which allows the AI to get all the information that the AST holds about a single symbol, including a parameter to include source code which completely replaces grepping and file reading for me.
The tool also specifies which other symbols call the one in question and which others it calls, respectively.
But yeah, sorry for this being already a quite long comment, if you want to give it a try, I published it on the VSCode marketplace a couple of days ago, and it's basically free right now, although I have to admit that I still want to try to earn a little bit of money with it at some point.
Right now, the daily usage limit is 2000 tool calls per day, which should be enough for anybody.
Would love to hear what you think :)
<https://marketplace.visualstudio.com/items?itemName=LuGoSoft...>
The fact that you've been using it for six months and that it performs well says a lot. At the end of the day, that's what counts.
I like your idea of piggybacking on top of the LSP services, and I can imagine that this was quite a bit of work. Doing it as an MCP server makes it usable across different tools.
I also really like the name "Context Master."
In my case, it's much more niche since it's for the tool I built. Though it's open source, the key difference is that the "indexing" is only agentic at this point.
I can see value in mixing the two. LSP integration scares me because of the amount of work involved, and tree-sitter seems like a good path.
In that case, in the code map, for each item, there could be both the LLM response info and some deterministic info, for example, from tree-sitter.
That being said, the current approach works so well that I think I am going to keep using and fine-tuning it for a while, and bring in deterministic context only when or if I need it.
Anyway, what you built looks great. If it works, that's great.
So, small model figures out which files to use based on the code map, and then enriches with snippets, so big model ideally gets preloaded with relevant context / snippets up front?
Where does code map live? Is it one big file?
I also have a `.tmpl-code-map.jsonl` in the same folder so all of my tasks can add to it, and then it gets merged into context-code-map.json.
I keep mtime, but I also compute a blake3 hash, so if mtime does not match, but it is just a "git restore," I do not redo the code map for that file. So it is very incremental.
Then the trick is, when sending the code map to AI, I serialize it in a nice, simple markdown format.
- path/to/file.rs - summary: ... - when to use: ... - public types: .., .., .. - public functions: .., .., ..
- ...
So the AI does not have to interpret JSON, just clean, structured markdown.
Funny, I worked on this addition to my tool for a week, planning everything, but even today, I am surprised by how well it works.
I have zero sed/grep in my workflow. Just this.
My prompt is pro@coder/coder-prompt.md, the first part is YAML for the globs, and the second part is my prompt.
There is a TUI, but all input and output are files, and the TUI is just there to run it and see the status.
1) Deterministic
2) Agentic So, this is why I started with #2.And then, the results in real coding scenarios have been astonishing.
Way above what I expected.
The way those indexes get combined with the user prompt gets the right files 95% of the time, and with surprisingly high quality.
So, I might add deterministic aspects to it, but since I think I will need the agentic step anyway, I have deprioritized it.
I imagine if the context were being commited and kept up-to-date with CI would work for others to use as well.
However, I'm a little confused on the autocontext/globs narrowing part. Do you, the developer, provide them? Or you feed the full code map to flash + your prompt so it returns the globs based on your prompt?
Also, in general, is your map of a file relatively smaller than the file itself, even for very small files?
- I have two main globs, which are lists of globs: knowledge_globs and context_globs. Knowledge can be absolute and should be relatively static. context_globs have to be relative to the workspace, since they are the working files.
- As a dev, you provide them in the top YAML section of the coder-prompt.md.
- The auto-context sub-agent calls the code-map sub-agent. Sub-agents can add to or narrow the given globs, and that is the goal of the auto-context agent.
It looks complicated, but it actually works like a charm.
Hopefully, I answered some of your questions.
I need to make a video about it.
But regardless, I really think it's not about the tools, it's about the techniques. This is where the true value is.
They're solving a different problem than you. So I think it's very plausible that you could come up with something that, for your use case, performs considerably better than their "defaults".
And that does not prevent mixing and matching the two, as some comments in this thread suggest.
Anyway, it's a great time for production coding.
> Standard pricing now applies across the full 1M window for both models, with no long-context premium. Media limits expand to 600 images or PDF pages.
For Claude Code users this is huge - assuming coherence remains strong past 200k tok.
No vibes allowed: https://youtu.be/rmvDxxNubIg?is=adMmmKdVxraYO2yQ
1) No longer found the dumb zone
2) No longer feared compaction
Switching to Opus for stupid political reasons, I still have not had the dumb zone - but I'm back to disliking compaction events and so the smaller context window it has, has really hurt.
I hope they copy OpenAI's compaction magic soon, but I am also very excited to try the longer context window.
His fix for "the dumb zone" is the RPI Framework:
● RESEARCH. Don't code yet. Let the agent scan the files first. Docs lie. Code doesn't.
● PLAN. The agent writes a detailed step-by-step plan. You review and approve the plan, not just the output. Dex calls this avoiding "outsourcing your thinking." The plan is where intent gets compressed before execution starts.
● IMPLEMENT. Execute in a fresh context window. The meta-principle he calls Frequent Intentional Compaction: don't let the chat run long. Ask the agent to summarize state, open a new chat with that summary, keep the model in the smart zone.
For me, it's less about being able to look back -800k tokens. It's about being able to flow a conversation for a lot longer without forcing compaction. Generally, I really only need the most recent ~50k tokens, but having the old context sitting around is helpful.
I'm using CC (Opus) thinking and Codex with xhigh on always.
And the models have gotten really good when you let them do stuff where goals are verifiable by the model. I had Codex fix a Rust B-rep CSG classification pipeline successfully over the course of a week, unsupervised. It had a custom STEP viewer that would take screenshots and feed them back into the model so it could verify the progress resp. the triangle soup (non progress) itself.
Codex did all the planning and verification, CC wrote the code.
This would have not been possible six months ago at all from my experience.
Maybe with a lot of handholding; but I doubt it (I tried).
I mean both the problem for starters (requires a lot of spatial reasoning and connected math) and the autonomous implementation. Context compression was never an issue in the entire session, for either model.
That said, 120k is pleeenty if you’re just building front-end components and have your API spec on hand already.
Dead Comment
I tried to ask questions about path of exile 2. And even with web research on it gave completely wrong information... Not only outdated. Wrong
I think context decay is a bigger problem then we feel like.
As an example of doing this in a session with jagged alliance 3 (an rpg) https://pastes.io/jagged-all-69136
Claude extracting game archives and dissasembling leads to far more reliable results than random internet posts.
It’s lead to me starting new chats with bigger and bigger starting ‘summary, prompts to catch the model up while refreshing it. Surely there’s a way to automate that technique.
The people I work with who complain about this type of thing horribly communicate their ask to the llm and expect it to read their minds.
In my experience the model will assume the web results are the answer even if the search engine returns irrelevant garbage.
For example you ask it a question about New Jersey law and the web results are about New York or about "many states" it'll assume the New York info or "many states" info is about New Jersey.
It'll remain a human job for quite a while too. Separability is not a property of vector spaces, so modern AIs are not going to be good at it. Maybe we can manage something similar with simplical complexes instead. Ideally you'd consult the large model once and say:
> show me the small contexts to use here, give me prompts re: their interfaces with their neighbors, and show me which distillations are best suited to those tasks
...and then a network of local models could handle it from there. But the providers have no incentive to go in that direction, so progress will likely be slow.
(Note that I'm using it in more of a hands-on pair-programming mode, and not in a fully-automated vibecoding mode.)
Deleted Comment
Dead Comment
I have not tested, but I would expect more niche ecosystems like Rust or Haskell or Erlang to have better overall training set (developer who care about good engineering focus on them), and potentially produce the best output.
For C and C++, I'd expect similar situation with Python: while not as approachable, it is also being pushed on beginning software engineers, and the training data would naturally have plenty of bad code.
Deleted Comment
Writing quick python scripts works a lot better than niche domain specific code
I have seen these shine on frontend work
Dead Comment
In practice, I haven't found this to be the case at all with Claude Code using Opus 4.6. So maybe it's another one of those things that used to be true, and now we all expect it to be true.
And of course when we expect something, we'll find it, so any mistakes at 150k context use get attributed to the context, while the same mistake at 50k gets attributed to the model.
Personally, I’m on a 6M+ line codebase and had no problems with the old window. I’m not sending it blindly into the codebase though like I do for small projects. Good prompts are necessary at scale.
I'm not an expert but maybe this explains context rot.
Normally buying the bigger plan gives some sort of discount.
At Claude, it's just "5 times more usage 5 times more cost, there you go".
They would probably implement _diminishing_-value pricing if pure pricing efficiency was their only concern.
Deleted Comment
EDIT: Don't think Pro has access to it, a typical prompt just hit the context limit.
The removal of extra pricing beyond 200k tokens may be Anthropic's salvo in the agent wars against GPT 5.4's 1M window and extra pricing for that.
I start with a PRD, ask for a step-by-step plan, and just execute on each step at a time. Sometimes ideas are dumb, but checking and guiding step by step helps it ship working things in hours.
It was also the first AI I felt, "Damn, this thing is smarter than me."
The other crazy thing is that with today's tech, these things can be made to work at 1k tokens/sec with multiple agents working at the same time, each at that speed.
It gave me an impressive plan of attack, including a reasonable way to determine which code it could safely modify. I told it to start with just a few files and let me review; its changes looked good. So I told it to proceed with the rest of the code.
It made hundreds of changes, as expected (big code base). And most of them were correct! Except the places where it decided to do things like put its "const x = useMemo(...)" call after some piece of code that used the value of "x", meaning I now had a bunch of undefined variable references. There were some other missteps too.
I tried to convince it to fix the places where it had messed up, but it quickly started wanting to make larger structural changes (extracting code into helper functions, etc.) rather than just moving the offending code a few lines higher in the source file. Eventually I gave up trying to steer it and, with the help of another dev on my team, fixed up all the broken code by hand.
It probably still saved time compared to making all the changes myself. But it was way more frustrating.
I have heard from people who regularly push a session through multiple compactions. I don’t think this is a good idea. I virtually never do this — when I see context getting up to even 100k, I start making sure I have enough written to disk to type /new, pipe it the diff so far, and just say “keep going.” I learned recently that even essentials like the CLAUDE.md part of the prompt get diluted through compactions. You can write a hook to re-insert it but it's not done by default.
This fresh context thing is a big reason subagents might work where a single agent fails. It’s not just about parallelism: each subagent starts with a fresh context, and the parent agent only sees the result of whatever the subagent does — its own context also remains clean.
There's probably a parallel with the CMSes and frameworks of the 2000s (e.g. WordPress or Ruby on Rails). They massively improved productivity, but as a junior developer you could get pretty stuck if something broke or you needed to implement an unconventional feature. I guess it must feel a bit similar for non-developers using tools like Claude Code today.
Definitely not ideal, but sure helps.
Deleted Comment
You need to converge on the requirements.
Just today I asked Claude using opus 4.6 to build out a test harness for a new dynamic database diff tool. Everything seemed to be fine but it built a test suite for an existing diff tool. It set everything up in the new directory, but it was actually testing code and logic from a preexisting directory despite the plan being correct before I told it to execute.
I started over and wrote out a few skeleton functions myself then asked it write tests for those to test for some new functionality. Then my plan was to the ask it to add that functionality using the tests as guardrails.
Well the tests didn’t actually call any of the functions under test. They just directly implemented the logic I asked for in the tests.
After $50 and 2 hours I finally got something working only to realize that instead of creating a new pg database to test against, it found a dev database I had lying around and started adding tables to it.
When I managed to fix that, it decided that it needed to rebuild multiple docker components before each test and test them down after each one.
After about 4 hours and $75, I managed to get something working that was probably more code than I would have written in 4 hours, but I think it was probably worse than what I would have come up with on my own. And I really have no idea if it works because the day was over and I didn’t have the energy left to review it all.
We’ve recently been tasked at work with spending more money on Claude (not being more productive the metric is literally spending more money) and everyone is struggling to do anything like what the posts on HN say they are doing. So far no one in my org in a very large tech company has managed to do anything very impressive with Claude other than bringing down prod 2 days ago.
Yes I’m using planning mode and clearing context and being specific with requirements and starting new sessions, and every other piece of advice I’ve read.
I’ve had much more luck using opus 4.6 in vs studio to make more targeted changes, explain things, debug etc… Claude seems too hard to wrangle and it isn’t good enough for you to be operating that far removed from the code.
I think it's the big investors' extremely powerful incentives manifesting in the form of internet comments. The pace of improvement peaked at GPT-4. There is value in autocomplete-as-a-service, and the "harnesses" like Codex take it a lot farther. But the people who are blown away by these new releases either don't spend a lot of time writing code, or are being paid to be blown away. This is not a hockey stick curve. It's a log curve.
Bigger context windows are a welcome addition. And stuff like JSON inputs is nice too. But these things aren't gonna like, take your SWE job, if you're any good. It's just like, a nice substitute for the Google -> Stack Overflow -> Copy/Paste workflow.
Huh? The max plan is $200/month. How are you spending $75 in 4 hrs?
- I ask for something highly general and claude explores a bit and responds.
- We go back and forth a bit on precisely what I'm asking for. Maybe I correct it a few times and maybe it has a few ideas I didn't know about/think of.
- It writes some kind of plan to a markdown file. In a fresh session I tell a new instance to execute the plan.
- After it's done, I skim the broad strokes of the code and point out any code/architectural smells.
- I ask it to review it's own work and then critique that review, etc. We write tests.
Perhaps that sounds like a lot but typically this process takes around 30-45 minutes of intermittent focus and the result will be several thousand lines of pretty good, working code.
GPT 5.4 on codex cli has been much more reliable for me lately. I used to have opus write and codex review, I now to the opposite (I actually have codex write and both review in parallel).
So on the latest models for my use case gpt > opus but these change all the time.
Edit: also the harness is shit. Claude code has been slow, weird and a resource hog. Refuses to read now standardized .agents dirs so I need symlink gymnastics. Hides as much info as it can… Codex cli is working much better lately.
Kinda funny how you don't actually need to use coercion if you put in the engineering work to build a product that's competitive on its own technical merits...
If you're not using AI you are cooked. You just don't realize it yet.
https://i.imgur.com/YXLZvy3.png
All programming is like this to some extent, but Claude's 80/20 behavior is so much more extreme. It can almost build anything in 15-30 minutes, but after those 15-30 minutes are up, it's only "almost built". Then you need to spend hours, days, maybe even weeks getting past the "almost".
Big part of why everyone seems to be vibe coding apps, but almost nobody seems to be shipping anything.
I had Opus 4.6 tell me I was "seeing things wrong" when I tried to have it correct some graphical issues. It got stuck in a loop of re-introducing the same bug every hour or so in an attempt to fix the issue.
I'm not disagreeing with your experience, but in my experience it is largely the same as what I had with Opus 4.5 / Codex / etc.
It started by insisting I was repeatedly making a typo and still would not budge even after I started copy/pasting the full terminal history of what I was entering and the unabridged output, and eventually pivoted to darkly insinuating I was tampering with my shell environment as if I was trying to mislead it or something.
Ultimately it turned out that it forgot it was supposed to be applying the fixes to the actual server instead of the local dev environment, and had earlier in the conversation switched from editing directly over SSH to pushing/pulling the local repo to the remote due to diffs getting mangled.
Deleted Comment
I also thought it was OPUS 4.5 (also tested a lot with 4.6) and then in February switched to only using auto mode in the coding IDEs. They do not use OPUS (most of the times), and I’m ending up with a similar result after a very rough learning curve.
Now switching back to OPUS I notice that I get more out of it, but it’s no longer a huge difference. In a lot of cases OPUS is actually in the way after learning to prompt more effectively with cheaper models.
The big difference now is that I’m just paying 60-90$ month for 40-50hrs of weekly usage… while I was inching towards 1000$ with OPUS. I chose these auto modes because they don’t dig into usage based pricing or throttling which is a pretty sweet deal.
Is it Baader-Meinhof or is everyone on HN suddenly using obscure acronyms?
Deleted Comment
Deleted Comment
It was about a problem with calculation around filling a topographical water basin with sedimentation where calculation is discrete (e.g. turn based) and that edge case where both water and sediments would overflow the basin; To make the matter simple, fact was A, B, C, and it oscillated between explanation 1 which refuted C, explanation 2 which refuted A and explanation 3 that refuted B.
I'll give it to opus training stability that my 3 tries using it all consistently got into this loop, so I decided to directly order it to do a brute force solution that avoided (but didn't solve) this problem.
I did feel like with a human, there's no way that those 3 loop would happen by the second time. Or at least the majority of us. But there is just no way to get through to opus 4.6
1000% agree. It's also easy to talk to it about something you're not sure it said and derive a better, more elegant solution with simple questioning.
Gemini 3.1 also gives me these vibes.
Horizontal parallelising of tasks doesn't really require any modern tech.
But I agree that Opus 4.6 with 1M context window is really good at lots of routine programming tasks.
Spent an hour or so unraveling the mess. My feeling are growing more and more conflicted about these tools. They are here to stay obviously.
I’m honestly uncertain about the junior engineers I’m working with who are more productive than they might be otherwise, but are gaining zero (or very little) experience. It’s like the future is a world where the entire programming sphere is dominated by the clueless non technical management that we’ve all had to deal with in small proportion a time or two.
Super simple problem :
I had a ZMK keyboard layout definition I wanted it to convert it to QMK for a different keyboard that had one key less so it just had to trim one outer key. It took like 45 minutes of back and forth to get it right - I could have done it in 30 min manually tops with looking up docs for everything.
Capability isn't the impressive part it's the tenacity/endurance.
Deleted Comment
That being said it's the only use case for me. I won't subscribe to something that I can't use with third party harness.
Not sure if this means I should get a more interesting job or if we are all going to be at the mercy of UBI eventually.
RIP widespread middle class. It was a good 80-year run.
Sounds like it is.
Also shout out to beads - I highly recommend you pair it with beads from yegge: opus can lay out a large project with beads, and keep track of what to do next and churn through the list beautifully with a little help.
The amount of genuine fuck-ups Codex finds makes me skeptical of people who are placing a lot of trust in Claude alone.
Not even close. There are still tons of architectural design issues that I'd find it completely useless at, tons of subtle issues it won't notice.
I never run agents by themselves; every single edit they do is approved by me. And, I've lost track of the innumerable times I've had to step in and redirect them (including Opus) to an objectively better approach. I probably should keep a log of all that, for the sake of posterity.
I'll grant you that for basic implementation of a detailed and well-specced design, it is capable.
Deleted Comment
I'm late to the party and I'm just getting started with Antrophic models. I have been finding Sonnet decent enough, but it seems to have trouble naming variables correctly (it's not just that most names are poor and undescriptive, sometimes it names it wrong, confusing) or sometimes unnecessarily declaring, re-declaring variables, encoding, decoding, rather than using the value that's already there etc. Is Opus better at this?
Dead Comment
Dead Comment
I'm working in a codebase of 200+ "microservices", separate repos, each deployed as multiple FaaS, CQRS-style. None of it my choice, everything precedes me, many repos I know nothing of. Little to no code re-use between them.
Any trace of "business logic" is so distributed in multiple repos, that I have no possible use of LLM codegen, unless I can somehow feed it ALL the codebase.
I've tried generating some tests, but they always miss the mark, as in the system under test is almost always the wrong one.
I guess LLM are cool for greenfield, but when the brownfield is really brown, there's no use for LLMs.
The stats claim Opus at 1M is about like 5.4 at 256k -- these needle long context tests don't always go with quality reasoning ability sadly -- but this is still a significant improvement, and I haven't seen dramatic falloff in my tests, unlike q4 '25 models.
p.s. what's up with sonnet 4.5 getting comparatively better as context got longer?
My employer only pays for GitHub copilot extension
https://github.com/features/copilot/cli
Disclosure: work at Msft