You can input any public GitHub repository, and it will automatically generate a README file for the codebase within 1-2 minutes.
I created this because I would often spend 1-2 hours writing up a perfect README file for each of my projects, and I thought generating them using an AI tool would make the process smoother.
I was inspired to create this project by GitIngest as well as the recent success of other tools like GitDiagram and DeepWiki."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Automatically generate a README file for your GitHub repository"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.gitread.dev/"}},"_tags":["story","author_vmath20","story_44007335","show_hn"],"author":"vmath20","created_at":"2025-05-16T16:30:32Z","created_at_i":1747413032,"num_comments":0,"objectID":"44007335","points":3,"story_id":44007335,"story_text":"GitRead is an open-source developer tool that utilizes recent advances in large context window LLMs to automatically generate README files.
You can input any public GitHub repository, and it will automatically generate a README file for the codebase within 1-2 minutes.
I created this because I would often spend 1-2 hours writing up a perfect README file for each of my projects, and I thought generating them using an AI tool would make the process smoother.
I was inspired to create this project by GitIngest as well as the recent success of other tools like GitDiagram and DeepWiki.","title":"Show HN: Automatically generate a README file for your GitHub repository","updated_at":"2025-05-19T09:03:30Z","url":"https://www.gitread.dev/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"MediaSquirrel"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"Transform any GitHub repo or folder on your Mac into a simple text file so that you can upload it to an LLM (Claude, ChatGPT, etc.) to reason over the whole thing in the context window.
LLMs like ChatGPT and Claude let you upload files, but they limit how many you can upload at once. When you're dealing with large codebases that have tons of files, you can't just upload them directly to an LLM. You end up having to use RAG (retrieval augmented generation) techniques, which in my experience aren't as effective as uploading everything into the full context window - especially when you need to reason about architecture or understand the entire system.
Flatty is simple to use.
Navigate to your project directory in the Terminal and then type "flatty" - that's it.
The script will output your GitHub repo as a text file (or series of text files in case of large repos) into a folder called "flattened" in your home directory.
I hope you find this tool useful in your own work!"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["llm"],"value":"Show HN: Flatty \u2013 Dead Simple Codebase-to-Text for LLMs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/mattmireles/Flatty"}},"_tags":["story","author_MediaSquirrel","story_42660498","show_hn"],"author":"MediaSquirrel","children":[42665823],"created_at":"2025-01-10T21:31:53Z","created_at_i":1736544713,"num_comments":1,"objectID":"42660498","points":1,"story_id":42660498,"story_text":"Transform any GitHub repo or folder on your Mac into a simple text file so that you can upload it to an LLM (Claude, ChatGPT, etc.) to reason over the whole thing in the context window.
LLMs like ChatGPT and Claude let you upload files, but they limit how many you can upload at once. When you're dealing with large codebases that have tons of files, you can't just upload them directly to an LLM. You end up having to use RAG (retrieval augmented generation) techniques, which in my experience aren't as effective as uploading everything into the full context window - especially when you need to reason about architecture or understand the entire system.
Flatty is simple to use.
Navigate to your project directory in the Terminal and then type "flatty" - that's it.
The script will output your GitHub repo as a text file (or series of text files in case of large repos) into a folder called "flattened" in your home directory.
I hope you find this tool useful in your own work!","title":"Show HN: Flatty \u2013 Dead Simple Codebase-to-Text for LLMs","updated_at":"2025-01-11T13:46:42Z","url":"https://github.com/mattmireles/Flatty"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"crush_robo_1536"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"https://rockyai.me/ now supports chatting with webpages that have content that exceeds the context window of LLM.
For example, on the recent HN: Who is hiring thread for Feb 2025[0], you can ask questions like "list product management roles posted here", "are there any machine learning engineer roles here", etc. You can also chat with pdfs on https://arxiv.org/ similarly without having to manually upload the pdf into chat gpt.
[0] https://news.ycombinator.com/item?id=42919502"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: In browser RAG on HN Who is hiring threads (and any other webpage)"}},"_tags":["story","author_crush_robo_1536","story_43075933","show_hn"],"author":"crush_robo_1536","created_at":"2025-02-17T06:43:27Z","created_at_i":1739774607,"num_comments":0,"objectID":"43075933","points":1,"story_id":43075933,"story_text":"https://rockyai.me/ now supports chatting with webpages that have content that exceeds the context window of LLM.
For example, on the recent HN: Who is hiring thread for Feb 2025[0], you can ask questions like "list product management roles posted here", "are there any machine learning engineer roles here", etc. You can also chat with pdfs on https://arxiv.org/ similarly without having to manually upload the pdf into chat gpt.
[0] https://news.ycombinator.com/item?id=42919502","title":"Show HN: In browser RAG on HN Who is hiring threads (and any other webpage)","updated_at":"2025-07-24T19:12:50Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"CShorten"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"Hey everyone! I am SUPER excited to publish our 73rd Weaviate Podcast with Charles Packer from UC Berkeley, the lead author of MemGPT!!
MemGPT is the Operating System for LLMs! This is one of the most exciting recent ideas in Large Language Model Applications that draws connections between Virtual Context Management in Operating Systems and the limited context windows of LLMs! One of the key ideas in MemGPT is to explicitly prompt the model that it has a limited token window and give it tools to self-edit its own context!!
This was such a fun discussion covering all things from the origin of the research, connections between operating systems and LLMs, details of the paper such as inference cost, explicit context annotation, recall and archival storage, and many more!
We also concluded the podcast with discussions around Creativity in AI and what directions from Multi-Agent Role Playing or Evolutionary methods will achieve truly creative, novel outputs from Generative AI systems!
I hope you enjoy the podcast! As always, more than happy to answer any questions or discuss any ideas about the content in the podcast!
https://www.youtube.com/watch?v=rxjsbUiuOFo"},"title":{"matchLevel":"none","matchedWords":[],"value":"Charles Packer on MemGPT"}},"_tags":["story","author_CShorten","story_38113392","ask_hn"],"author":"CShorten","created_at":"2023-11-02T13:42:30Z","created_at_i":1698932550,"num_comments":0,"objectID":"38113392","points":1,"story_id":38113392,"story_text":"Hey everyone! I am SUPER excited to publish our 73rd Weaviate Podcast with Charles Packer from UC Berkeley, the lead author of MemGPT!!
MemGPT is the Operating System for LLMs! This is one of the most exciting recent ideas in Large Language Model Applications that draws connections between Virtual Context Management in Operating Systems and the limited context windows of LLMs! One of the key ideas in MemGPT is to explicitly prompt the model that it has a limited token window and give it tools to self-edit its own context!!
This was such a fun discussion covering all things from the origin of the research, connections between operating systems and LLMs, details of the paper such as inference cost, explicit context annotation, recall and archival storage, and many more!
We also concluded the podcast with discussions around Creativity in AI and what directions from Multi-Agent Role Playing or Evolutionary methods will achieve truly creative, novel outputs from Generative AI systems!
I hope you enjoy the podcast! As always, more than happy to answer any questions or discuss any ideas about the content in the podcast!
https://www.youtube.com/watch?v=rxjsbUiuOFo","title":"Charles Packer on MemGPT","updated_at":"2024-09-20T15:32:54Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ragall"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"You're wrong on every count: given the small context window of LLMs, the need for interfaces and scehmas is even greater."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Protobuf has LSP support"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://buf.build/blog/protobuf-lsp"}},"_tags":["comment","author_ragall","story_49322573"],"author":"ragall","comment_text":"You're wrong on every count: given the small context window of LLMs, the need for interfaces and scehmas is even greater.","created_at":"2026-08-17T05:34:26Z","created_at_i":1786944866,"objectID":"49326885","parent_id":49322885,"story_id":49322573,"story_title":"Protobuf has LSP support","story_url":"https://buf.build/blog/protobuf-lsp","updated_at":"2026-08-17T14:05:31Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Aurornis"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"AI completely broke some of the signals management used as proxies for team competence. In the past, teams that moved quickly and had a proof of concept to show management were valued for the speed of their execution. Usually a team that could execute quickly had some competent developers who could also quickly execute on the challenges of the implementation when it scaled up to real world problems.
Now anyone with enough LLM tokens can move quickly and have a pretty proof of concept in front of management in days or maybe hours. There's a land grab to get things in front of management so you can stake your claim on something. The implementation details are assumed to be solvable by enough Claude sessions and ChatGPT prompts.
Where this falls apart is that the real difficult problems don't scale well with LLM brute force techniques. You need people who understand the problem and know how to drive the LLM agents around properly, as well as when to set them aside and go look at the code and data themselves.
The teams who try to brute force everything with LLM tokens can get a little farther every few months, but for complex problems I've seen enough variations of teams hitting a wall of complexity and tech debt with a team that doesn't know much about how their system works because they've just been asking agents to solve every problem as it comes up. With each iteration it takes a little longer for the agents to come to a solution and the system becomes a little more brittle. Even with our current mega-context-window LLMs you can reach a point where every other change to your big ball of code breaks something else, which requires another agent session to diagnose and fix, which might then break another thing down the line. The refactors to try to clean things up end up breaking even more. Having a zillion tests starts to backfire when the team can't tell which tests are testing valid results and which are baking in incorrect results.
Maybe there's an end game where agents get so good that none of this matters, but I think it's more likely that we're going to have to build up better intuitions about what good execution looks like. The old signals of moving fast are meaningly and potentially even negative now."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"The New AI Superpowers: Focus and Followthrough"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.rickmanelius.com/p/the-new-ai-superpowers-focus-and"}},"_tags":["comment","author_Aurornis","story_49057877"],"author":"Aurornis","children":[49065621,49066162,49068996],"comment_text":"AI completely broke some of the signals management used as proxies for team competence. In the past, teams that moved quickly and had a proof of concept to show management were valued for the speed of their execution. Usually a team that could execute quickly had some competent developers who could also quickly execute on the challenges of the implementation when it scaled up to real world problems.
Now anyone with enough LLM tokens can move quickly and have a pretty proof of concept in front of management in days or maybe hours. There's a land grab to get things in front of management so you can stake your claim on something. The implementation details are assumed to be solvable by enough Claude sessions and ChatGPT prompts.
Where this falls apart is that the real difficult problems don't scale well with LLM brute force techniques. You need people who understand the problem and know how to drive the LLM agents around properly, as well as when to set them aside and go look at the code and data themselves.
The teams who try to brute force everything with LLM tokens can get a little farther every few months, but for complex problems I've seen enough variations of teams hitting a wall of complexity and tech debt with a team that doesn't know much about how their system works because they've just been asking agents to solve every problem as it comes up. With each iteration it takes a little longer for the agents to come to a solution and the system becomes a little more brittle. Even with our current mega-context-window LLMs you can reach a point where every other change to your big ball of code breaks something else, which requires another agent session to diagnose and fix, which might then break another thing down the line. The refactors to try to clean things up end up breaking even more. Having a zillion tests starts to backfire when the team can't tell which tests are testing valid results and which are baking in incorrect results.
Maybe there's an end game where agents get so good that none of this matters, but I think it's more likely that we're going to have to build up better intuitions about what good execution looks like. The old signals of moving fast are meaningly and potentially even negative now.","created_at":"2026-07-27T05:38:20Z","created_at_i":1785130700,"objectID":"49065530","parent_id":49061112,"story_id":49057877,"story_title":"The New AI Superpowers: Focus and Followthrough","story_url":"https://www.rickmanelius.com/p/the-new-ai-superpowers-focus-and","updated_at":"2026-07-28T00:22:20Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"SyneRyder"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":">Problem 3: "you'll hit the 200k token limit..." ... Suggestion: use 1 million context window LLMs.
Yes, if the model someone is using only has 200k token limit, that would immediately suggest to me that it really isn't a sophisticated enough model.
Most of my coding sessions end up being about 350k tokens long when I finish, it wouldn't even fit in a 200k context. And that isn't counting the cache-reads by subagents, etc.
It's worth spending some time with the best Opus / GPT model, to at least get a sense of what the frontier is like."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"AI coding is a nightmare. Am I the only one experiencing this?"}},"_tags":["comment","author_SyneRyder","story_48770319"],"author":"SyneRyder","children":[48773221],"comment_text":">Problem 3: "you'll hit the 200k token limit..." ... Suggestion: use 1 million context window LLMs.
Yes, if the model someone is using only has 200k token limit, that would immediately suggest to me that it really isn't a sophisticated enough model.
Most of my coding sessions end up being about 350k tokens long when I finish, it wouldn't even fit in a 200k context. And that isn't counting the cache-reads by subagents, etc.
It's worth spending some time with the best Opus / GPT model, to at least get a sense of what the frontier is like.","created_at":"2026-07-03T09:56:13Z","created_at_i":1783072573,"objectID":"48773101","parent_id":48772685,"story_id":48770319,"story_title":"AI coding is a nightmare. Am I the only one experiencing this?","updated_at":"2026-07-03T12:09:22Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"bel8"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"Yes I've experienced everything you stated. Here's what helped me:
Problem 1: "Obsessed with reinventing the wheel" " three duplicate functions":
Suggestion: plan then implement.
Tell LLM to scan your project and crete markdown file plan to solve the task first. DO NOT try to selve tasks in a single shot without planning. Review the plan file then, IN A NEW SESSION with clean context, tell LLM to read the implementation plan file and implement the plan according to the file.
---
Problem 2: "hyper-focuses on the current task and couldn't care less if its changes break other parts of the system"
Suggestion: add instructions to AGENTS.md file teaching LLMs how to run unit tests and other kinds of tests so it can make sure nothing broke. And also add to AGENTS.md that LLMs MUST run tests before marking the task done.
---
Problem 3: "you'll hit the 200k token limit in no time" "Long context = instant brain damage"
Suggestion: use 1 million context window LLMs. Also plan then implement will keep your context shorter.
If you can, use better LLM services which offer 1million context window. If you can't afford Anthropic or OpenAI, use DeepSeek V4 Flash or MiMo 2.5 for example. A $10/mo OpenCode Go subscription plan offers $60 in LLM credits which is A LOT for these cheap LLMs.
Also, planning phase is when the LLM has to scan the entire project to understand what needs changing. This is where the context bloat comes from. If you split tasks into planning + implementation, the scanning phase is condensed into a single markdown file which keeps context lean.
Bonus tip: Tell LLMs to use subagents when doing exploration.
---
Problem 4: The longer the context, the more incoherent its responses.
Suggestion: yeah, LLMs get dumber as their working memory fills up (just like me). If your session reaches 200k+ tokens, it's usually a sign you could have planned the feature better or split it up. It might be worth restarting with more clarification."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"AI coding is a nightmare. Am I the only one experiencing this?"}},"_tags":["comment","author_bel8","story_48770319"],"author":"bel8","children":[48773101],"comment_text":"Yes I've experienced everything you stated. Here's what helped me:
Problem 1: "Obsessed with reinventing the wheel" " three duplicate functions":
Suggestion: plan then implement.
Tell LLM to scan your project and crete markdown file plan to solve the task first. DO NOT try to selve tasks in a single shot without planning. Review the plan file then, IN A NEW SESSION with clean context, tell LLM to read the implementation plan file and implement the plan according to the file.
---
Problem 2: "hyper-focuses on the current task and couldn't care less if its changes break other parts of the system"
Suggestion: add instructions to AGENTS.md file teaching LLMs how to run unit tests and other kinds of tests so it can make sure nothing broke. And also add to AGENTS.md that LLMs MUST run tests before marking the task done.
---
Problem 3: "you'll hit the 200k token limit in no time" "Long context = instant brain damage"
Suggestion: use 1 million context window LLMs. Also plan then implement will keep your context shorter.
If you can, use better LLM services which offer 1million context window. If you can't afford Anthropic or OpenAI, use DeepSeek V4 Flash or MiMo 2.5 for example. A $10/mo OpenCode Go subscription plan offers $60 in LLM credits which is A LOT for these cheap LLMs.
Also, planning phase is when the LLM has to scan the entire project to understand what needs changing. This is where the context bloat comes from. If you split tasks into planning + implementation, the scanning phase is condensed into a single markdown file which keeps context lean.
Bonus tip: Tell LLMs to use subagents when doing exploration.
---
Problem 4: The longer the context, the more incoherent its responses.
Suggestion: yeah, LLMs get dumber as their working memory fills up (just like me). If your session reaches 200k+ tokens, it's usually a sign you could have planned the feature better or split it up. It might be worth restarting with more clarification.","created_at":"2026-07-03T09:04:47Z","created_at_i":1783069487,"objectID":"48772685","parent_id":48770319,"story_id":48770319,"story_title":"AI coding is a nightmare. Am I the only one experiencing this?","updated_at":"2026-07-03T21:39:09Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jstanley"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"The most interesting thing in here is https://github.com/smhanov/laconic which is the author's "agentic research orchestrator for Go that is optimized to use free search & low-cost limited context window llms".
I have been doing this kind of thing with Cursor and Codex subscriptions, but they do have annoying rate limits, and Cursor on the Auto model seems to perform poorly if you ask it to do too much work, so I am keen to try out laconic on my local GPU.
EDIT:
Having tried it out, this may be a false economy.
The way it works is it has a bunch of different prompts for the LLMs (Planner, Synthesizer, Finalizer).
The "Planner" is given your input question and the "scratchpad" and has to come up with DuckDuckGo search terms.
Then the harness runs the DuckDuckGo search and gives the question, results, and scratchpad to the Synthesizer. The Synthesizer updates the scratchpad with new information that is learnt.
This continues in a loop, with the Planner coming up with new search queries and the Synthesizer updating the scratchpad, until eventually the Planner decides to give a final answer, at which point the Finalizer summarises the information in a user-friendly final answer.
That is a pretty clever design! It allows you to do relatively complex research with only a very small amount of context window. So I love that.
However I have found that the Synthesizer step is extremely slow on my RTX3060, and also I think it would cost me about \u00a31/day extra to run the RTX3060 flat out vs idle. For the amount of work laconic can do in a day (not a lot!), I think I am better off just sending the money to OpenAI and getting the results more quickly.
But I still love the design, this is a very creative way to use a very small context window. And has the obvious privacy and freedom advantages over depending on OpenAI."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"I run multiple $10K MRR companies on a $20/month tech stack"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://stevehanov.ca/blog/how-i-run-multiple-10k-mrr-companies-on-a-20month-tech-stack"}},"_tags":["comment","author_jstanley","story_47736555"],"author":"jstanley","children":[47737616],"comment_text":"The most interesting thing in here is https://github.com/smhanov/laconic which is the author's "agentic research orchestrator for Go that is optimized to use free search & low-cost limited context window llms".
I have been doing this kind of thing with Cursor and Codex subscriptions, but they do have annoying rate limits, and Cursor on the Auto model seems to perform poorly if you ask it to do too much work, so I am keen to try out laconic on my local GPU.
EDIT:
Having tried it out, this may be a false economy.
The way it works is it has a bunch of different prompts for the LLMs (Planner, Synthesizer, Finalizer).
The "Planner" is given your input question and the "scratchpad" and has to come up with DuckDuckGo search terms.
Then the harness runs the DuckDuckGo search and gives the question, results, and scratchpad to the Synthesizer. The Synthesizer updates the scratchpad with new information that is learnt.
This continues in a loop, with the Planner coming up with new search queries and the Synthesizer updating the scratchpad, until eventually the Planner decides to give a final answer, at which point the Finalizer summarises the information in a user-friendly final answer.
That is a pretty clever design! It allows you to do relatively complex research with only a very small amount of context window. So I love that.
However I have found that the Synthesizer step is extremely slow on my RTX3060, and also I think it would cost me about \u00a31/day extra to run the RTX3060 flat out vs idle. For the amount of work laconic can do in a day (not a lot!), I think I am better off just sending the money to OpenAI and getting the results more quickly.
But I still love the design, this is a very creative way to use a very small context window. And has the obvious privacy and freedom advantages over depending on OpenAI.","created_at":"2026-04-12T08:09:12Z","created_at_i":1775981352,"objectID":"47737193","parent_id":47736555,"story_id":47736555,"story_title":"I run multiple $10K MRR companies on a $20/month tech stack","story_url":"https://stevehanov.ca/blog/how-i-run-multiple-10k-mrr-companies-on-a-20month-tech-stack","updated_at":"2026-04-12T16:49:30Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"loehnsberg"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"I think if we want to build on what we have, instead of compaction at the end of the context window, the LLM would have to 'sleep', i.e. adjust its weights, then wake up with the last bits of the old context window in the new one, and have a 'feel' for what it did before through the change in weights. I just sense it's not that simple to get there, because simply updating the weights based on a single context sample risks degrading the weights of the whole network.
I like the idea of using small local model (or several) for tackling this problem, like low rank adaptation, but with current tech, I still have to piece this together or the small local models will forget old memories."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"OpenClaw\u2019s memory is unreliable, and you don\u2019t know when it will break"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://blog.nishantsoni.com/p/ive-seen-a-thousand-openclaw-deploys"}},"_tags":["comment","author_loehnsberg","story_47721955"],"author":"loehnsberg","children":[47729193,47733121],"comment_text":"I think if we want to build on what we have, instead of compaction at the end of the context window, the LLM would have to 'sleep', i.e. adjust its weights, then wake up with the last bits of the old context window in the new one, and have a 'feel' for what it did before through the change in weights. I just sense it's not that simple to get there, because simply updating the weights based on a single context sample risks degrading the weights of the whole network.
I like the idea of using small local model (or several) for tackling this problem, like low rank adaptation, but with current tech, I still have to piece this together or the small local models will forget old memories.","created_at":"2026-04-11T09:17:30Z","created_at_i":1775899050,"objectID":"47728946","parent_id":47728494,"story_id":47721955,"story_title":"OpenClaw\u2019s memory is unreliable, and you don\u2019t know when it will break","story_url":"https://blog.nishantsoni.com/p/ive-seen-a-thousand-openclaw-deploys","updated_at":"2026-04-11T21:46:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"andychen32"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"I agree! my instincts tell me most enterprises will hop onto this in the coming months (late stave startups first, then bigger companies)\nuntil we get to 100m token context window LLMs, this is probably the final pattern"},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["context"],"value":"The Enterprise Context Layer"},"story_url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["context"],"value":"https://andychen32.substack.com/p/the-enterprise-context-layer"}},"_tags":["comment","author_andychen32","story_47324448"],"author":"andychen32","comment_text":"I agree! my instincts tell me most enterprises will hop onto this in the coming months (late stave startups first, then bigger companies)\nuntil we get to 100m token context window LLMs, this is probably the final pattern","created_at":"2026-03-17T15:01:24Z","created_at_i":1773759684,"objectID":"47413659","parent_id":47327942,"story_id":47324448,"story_title":"The Enterprise Context Layer","story_url":"https://andychen32.substack.com/p/the-enterprise-context-layer","updated_at":"2026-05-22T19:41:13Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rirze"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"At this point, I feel like MCP servers are just not feasible at the current level of context windows and LLMs. Good idea, but we're way too early."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["context"],"value":"Apideck CLI \u2013 An AI-agent interface with much lower context consumption than MCP"},"story_url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["context","window"],"value":"https://www.apideck.com/blog/mcp-server-eating-context-window-cli-alternative"}},"_tags":["comment","author_rirze","story_47400261"],"author":"rirze","comment_text":"At this point, I feel like MCP servers are just not feasible at the current level of context windows and LLMs. Good idea, but we're way too early.","created_at":"2026-03-16T15:45:44Z","created_at_i":1773675944,"objectID":"47400549","parent_id":47400261,"story_id":47400261,"story_title":"Apideck CLI \u2013 An AI-agent interface with much lower context consumption than MCP","story_url":"https://www.apideck.com/blog/mcp-server-eating-context-window-cli-alternative","updated_at":"2026-03-17T18:06:28Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"freehorse"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"No, I don't trust an encyclopedia generated by AI. Projects with much narrower scopes are not comparable.
edit: I am not very excited by AI-generated documentations either. I think that LLMs are very useful tools, but I see a potential problem when the sources of information that their usefulness is largely based on is also LLM-generated. I am afraid that this will inevitably result in drop in quality that will also affect the LLMs themselves downstream. I think we underestimate the importance that intentionality in human-written text plays in being in the training sets/context windows of LLMs for them to give relevant/useful output."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Elon Musk pushes out more xAI founders as AI coding effort falters"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.ft.com/content/e5fbc6c2-d5a6-4b97-a105-6a96ea849de5"}},"_tags":["comment","author_freehorse","story_47366666"],"author":"freehorse","comment_text":"No, I don't trust an encyclopedia generated by AI. Projects with much narrower scopes are not comparable.
edit: I am not very excited by AI-generated documentations either. I think that LLMs are very useful tools, but I see a potential problem when the sources of information that their usefulness is largely based on is also LLM-generated. I am afraid that this will inevitably result in drop in quality that will also affect the LLMs themselves downstream. I think we underestimate the importance that intentionality in human-written text plays in being in the training sets/context windows of LLMs for them to give relevant/useful output.","created_at":"2026-03-13T22:12:57Z","created_at_i":1773439977,"objectID":"47370667","parent_id":47370494,"story_id":47366666,"story_title":"Elon Musk pushes out more xAI founders as AI coding effort falters","story_url":"https://www.ft.com/content/e5fbc6c2-d5a6-4b97-a105-6a96ea849de5","updated_at":"2026-03-24T09:26:09Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"reverius42"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"Not sure that matters when the entire language spec is expected to be part of every context window? The LLM should be capable of using the language spec to express intended precedence."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: The Mog Programming Language"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://moglang.org"}},"_tags":["comment","author_reverius42","story_47312728"],"author":"reverius42","comment_text":"Not sure that matters when the entire language spec is expected to be part of every context window? The LLM should be capable of using the language spec to express intended precedence.","created_at":"2026-03-09T23:31:04Z","created_at_i":1773099064,"objectID":"47317214","parent_id":47315306,"story_id":47312728,"story_title":"Show HN: The Mog Programming Language","story_url":"https://moglang.org","updated_at":"2026-03-09T23:32:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vb7132"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"
IMO, the writer is overzealous with their comments on LLMs. As a coder, it feels like an outsider trying out a product that was amazed me over and over so many times.\n\n > They aren\u2019t perfect, but the kind of analysis the program is able to do is past the point where technology looks like magic.\n\n But as you use this product over a long period of time, there are many obvious gaps - hallucinations / repeated tool calls / out of context outputs / etc.\n\n To me, refine.ink sounds like a company that has built heavy tooling around some super high context window LLMs and then some very good prompts. Their claim is to compare it against any good off-the-shelf LLM with any prompt. But when you are spending bunch of money to build a whole ecosystem around LLMs, it's obvious that it's not going to beat their output. \n\n I won't be surprised if the next version of an LLM within the next few months completely outperforms their output -- that's usually the case with all the coding tools and scaffoldings. They are rendered useless by a superior LLM. IMO, the writer is overzealous with their comments on LLMs. As a coder, it feels like an outsider trying out a product that was amazed me over and over so many times.\n\n > They aren\u2019t perfect, but the kind of analysis the program is able to do is past the point where technology looks like magic.\n\n But as you use this product over a long period of time, there are many obvious gaps - hallucinations / repeated tool calls / out of context outputs / etc.\n\n To me, refine.ink sounds like a company that has built heavy tooling around some super high context window LLMs and then some very good prompts. Their claim is to compare it against any good off-the-shelf LLM with any prompt. But when you are spending bunch of money to build a whole ecosystem around LLMs, it's obvious that it's not going to beat their output. \n\n I won't be surprised if the next version of an LLM within the next few months completely outperforms their output -- that's usually the case with all the coding tools and scaffoldings. They are rendered useless by a superior LLM.Beware I'm a complete AI layman. All this is from background reading of popular articles. It may well be wrong. It's definitely out of date.
It has to do with how the attention heads work. The attention heads (the idea originated from the "Attention is all you need" paper, arguably the single most important AI paper to date), direct the LLM to work on the most relevant parts of the conversation. If you want a human analogue, it's your attention heads that are tacking the interesting points in a conversation.
The original attention heads output a relevance score for every pair of words in the context window. Thus in "Time flies like an arrow", it's the attention heads that spot the word "Time" is very relevant to "arrow", but not "flies". The implication of this is an attention head does O(N*N) work. It does not scale well to large context windows.
Nonetheless, you see claims of "large" context windows the LLMs marketing. (Large is in quotes, because even a 1M context window begins to feel very cramped in a write / test / fix loop.) But a 1M context-window would require a attention head requiring a 1 trillion element matrix. That isn't feasible. The industry even has a name for the size of the window they give in their marketing: the Effective Context Window. Internally they have another metric that measures the real amount of compute they throw at attention: the Physical Context Window. The bridge between the two is some proprietary magic that discards tokens in the context window that are likely to be irrelevant. In my experience, that bridge is pretty good at doing that, where "pretty good" is up to human standards.
But eventually (actually quickly in my experience), you fill up even the marketed size of the context window because it is remembering every word said, in the order they were said. If it reads code it's written to debug it, it appears twice in the context window. All compiler and test output also ends up there. Once the context window fills up they take drastic action, because it like letting malloc fail. Even reporting a malloc failure is hard because it usually needs more malloc to do the reporting. Anthropic calls it compacting. It throws away 90% of your tokens. It turns your helpful LLM into a goldfish with dementia. It is nowhere near as good as human is at remembering what happened. Not even close."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["llm"],"value":"LLM=True"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://blog.codemine.be/posts/2026/20260222-be-quiet/"}},"_tags":["comment","author_rstuart4133","story_47149151"],"author":"rstuart4133","comment_text":"> I think my question at this point is what about this is specific to LLMs. Humans should not be forced to wade through reams of garbage output either.
Beware I'm a complete AI layman. All this is from background reading of popular articles. It may well be wrong. It's definitely out of date.
It has to do with how the attention heads work. The attention heads (the idea originated from the "Attention is all you need" paper, arguably the single most important AI paper to date), direct the LLM to work on the most relevant parts of the conversation. If you want a human analogue, it's your attention heads that are tacking the interesting points in a conversation.
The original attention heads output a relevance score for every pair of words in the context window. Thus in "Time flies like an arrow", it's the attention heads that spot the word "Time" is very relevant to "arrow", but not "flies". The implication of this is an attention head does O(N*N) work. It does not scale well to large context windows.
Nonetheless, you see claims of "large" context windows the LLMs marketing. (Large is in quotes, because even a 1M context window begins to feel very cramped in a write / test / fix loop.) But a 1M context-window would require a attention head requiring a 1 trillion element matrix. That isn't feasible. The industry even has a name for the size of the window they give in their marketing: the Effective Context Window. Internally they have another metric that measures the real amount of compute they throw at attention: the Physical Context Window. The bridge between the two is some proprietary magic that discards tokens in the context window that are likely to be irrelevant. In my experience, that bridge is pretty good at doing that, where "pretty good" is up to human standards.
But eventually (actually quickly in my experience), you fill up even the marketed size of the context window because it is remembering every word said, in the order they were said. If it reads code it's written to debug it, it appears twice in the context window. All compiler and test output also ends up there. Once the context window fills up they take drastic action, because it like letting malloc fail. Even reporting a malloc failure is hard because it usually needs more malloc to do the reporting. Anthropic calls it compacting. It throws away 90% of your tokens. It turns your helpful LLM into a goldfish with dementia. It is nowhere near as good as human is at remembering what happened. Not even close.","created_at":"2026-02-26T02:25:57Z","created_at_i":1772072757,"objectID":"47161027","parent_id":47150870,"story_id":47149151,"story_title":"LLM=True","story_url":"https://blog.codemine.be/posts/2026/20260222-be-quiet/","updated_at":"2026-03-05T23:37:34Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"burnerToBetOut"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["context","window","llm"],"value":"The answer, my friendow, is blowing in the context window [1]\u2026
_______
\u2026LLMs already "know" these books deeply, but without a structured prompt scaffold they apply that knowledge inconsistently and at low confidence. Giving the model a explicit lens \u2014 "review this as if you're checking against Clean Code heuristics C1\u2013C36" \u2014 concentrates attention and dramatically reduces hallucinated or off-topic feedback\u2026
\u2026
Where I'd push back or warn you:
Context collapse is your #1 enemy. Clean Code was written for Java in 2008. DDIA is about distributed systems at scale. If you apply the Clean Code reviewer to a 50-line Python script, you'll get pedantic nonsense about function length when the actual problem might be that the data model is wrong. Your skill selection logic needs to be domain-aware, not just "throw all skills at every file"\u2026
\u2026
_______
[1] https://g2ww.short.gy/ZLStasQ1"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Using classic dev books to guide AI agents"}},"_tags":["comment","author_burnerToBetOut","story_47098555"],"author":"burnerToBetOut","comment_text":"The answer, my friendow, is blowing in the context window [1]\u2026
_______
\u2026LLMs already "know" these books deeply, but without a structured prompt scaffold they apply that knowledge inconsistently and at low confidence. Giving the model a explicit lens \u2014 "review this as if you're checking against Clean Code heuristics C1\u2013C36" \u2014 concentrates attention and dramatically reduces hallucinated or off-topic feedback\u2026
\u2026
Where I'd push back or warn you:
Context collapse is your #1 enemy. Clean Code was written for Java in 2008. DDIA is about distributed systems at scale. If you apply the Clean Code reviewer to a 50-line Python script, you'll get pedantic nonsense about function length when the actual problem might be that the data model is wrong. Your skill selection logic needs to be domain-aware, not just "throw all skills at every file"\u2026
\u2026
_______
[1] https://g2ww.short.gy/ZLStasQ1","created_at":"2026-02-21T11:01:04Z","created_at_i":1771671664,"objectID":"47099583","parent_id":47098555,"story_id":47098555,"story_title":"Show HN: Using classic dev books to guide AI agents","updated_at":"2026-03-05T23:37:00Z"}],"hitsPerPage":20,"nbHits":4776,"nbPages":50,"page":0,"params":"query=context+window+LLM&advancedSyntax=true&analyticsTags=backend","processingTimeMS":32,"processingTimingsMS":{"_request":{"roundTrip":14},"afterFetch":{"format":{"highlighting":1,"total":1},"merge":{"mergeLoop":{"prepareNextHit":4,"total":4},"total":5},"total":5},"fetch":{"query":6,"scanning":18,"total":25},"total":32},"query":"context window LLM","serverTimeMS":33}