I will try to write it myself and update it, thanks for feedback!
Example: https://yazzy.carter.works/https://paulgraham.com/submarine....
https://github.com/carterworks/yazzy
It uses Steph Ango's/kepano's defuddle under the hood.
The live version is hosted on a single free-tier fly.io node, but it is easily self-hosted.
Or perhaps the LLM should have known to do that.
I am dyslexic so have a hard time arranging things, so I gave the basic idea and the AI generated the README
They claim the protocol is resilient to enshittification.
It does seem like a lot of computational effort to achieve what F9 / Reader View does in FF.
I tried several reader modes, there were several issues including * several potions of the main content was missing * the navigation bits get caught when in reader mode * the comments and other un-related sections come in play
I really tried these before invesitng time in this
Also, I understand writing READMEs is boring, but please at least edit what the LLM produces. You do not need this many content-free emoji bullet point sections.
Edit: Looking at the prompt made me realize that the output of this would obviously be completely untrustworthy: https://github.com/subranag/declutter/blob/main/src/llm.ts#L...
why is everyone against emojis :) I personally feel they are fun!
I Am dyslexic so I have a bunch of ideas and I have a hard time laying that out in coherent text, so I gave the basic idea to the LLM and it wrote the README in the structure I want.
but feedback taken I will try to write it myself, thanks for taking time out to provide feedback!
Sorry, let me ask ChatGPT to put it in terms people seem to prefer now (I don't think this stuff is actually quite right but who cares anymore):
## 1. They Optimize for Politeness, Not Usefulness
ChatGPT READMEs tend to:
- Over-explain obvious things
- Avoid strong claims
- Hedge unnecessarily
The result is text that feels safe but not informative. A good README should reduce uncertainty quickly, not pad it with disclaimers and filler.
## 2. They Follow Templates Instead of Intent
Most generated READMEs look structurally correct but contextually shallow:
- Generic section headings (“Installation”, “Usage”, “Contributing”) regardless of relevance
- Boilerplate language that could apply to almost any project
- No clear prioritization of what actually matters
This signals that the README was assembled, not written with purpose.
## Summary
ChatGPT READMEs are usually:
- Correct but unhelpful
- Polished but shallow
- Complete but low-signalI will take your feedback and see if I can restructure it better! thanks for taking time out to structure your comment properly, gives me a good insight on how to write good READMEs
The modern web is broken. Before you can read anything, you hit a wall: popups,
ads, paywalls, tracking scripts, and navigation clutter designed to keep you clicking
instead of reading.
Declutter strips all of it away using AI. But here's the thing, you don't need
frontier models. Even small, cheap models like Gemini Flash or Claude Haiku do an
excellent job extracting pure content. That means you can archive whatever you want
to read for pennies per month.
Just point it at a URL, and out comes beautifully formatted Markdown, HTML, or PDF.
Offline. Local. Clean.
I built this because I was tired of the fight. The web doesn't have to be this way.
I just want to read. Please let me read without distractions.