This is one of the worst pieces of unedited LLM slop I've seen yet. At least read what it outputs instead of doing 1 sentence prompt -> ctrl+c ctrl+v -> publish...
This is clickbait speak. Stop talking like this or titling things like this. Dump what you are trying to say immediately, and put some flavor in it. Example
“LLMs.txt and MCP are useless pieces of shit, bad ideas.”
Also the subsequent paragraphs. It’s laid out like basic ChatGPT answers for the last year.
Statement: basic wordy elaboration.
Statement: basic wordy elaboration.
If this was actually a person writing, which I don’t believe it was, then they are using way too much AI and are actually being negatively impacted. They’re mimicking the tool that mimics us.
"move the needle" is a Claudism, sure, but the text itself doesn't pop out at me. There are spelling mistakes and abbreviations like "tbh". It's clearly at least been edited.
Does anyone have any empirical evidence that llms.txt makes any difference?
I've had multiple marketeers try to convince me that it's a must for a website, but the best "proof" they could give me were some LinkedIn posts from other marketing consultants.
I've tried replicating a scenario where Claude/ChatGPT would look for this file, but without luck.
It's just a token saving mechanism. My llms.txt are a text index/sitemap for the agent. AFAIK Claude doesn't look for this file, it needs to be explicitly instructed to.
Fair, I still find SOTA models insufficient in this regard. Usually they do not have a comprehensive enough grasp of the the complete feature set or latest changes, in my experience. I find context7 a bit more helpful, tho I sometimes just make my own docs of libraries I use too so the agent just loads my compiled docs versions for the libraries I want expertise in instead. You're prly right tho that this could likely be solved through better prompting of the SOTA models to be much more comprehensive in their understanding, to avoid reading just a few pages and thinking they have a complete understanding, which they still often do, even SOTA.
Only one data point, but I get a lot of submissions for my newsletter from agents. Months ago, I updated llms.txt to include specific instructions for submissions that differ from what you can figure out from the human site, and I haven't seen a single one follow the rules in llms.txt yet. I can't say I've noticed any of my own agents looking for them either when just reading normal pages is a bread-and-butter task for them now.
antiloper | 9 hours ago
StilesCrisis | 9 hours ago
deadbabe | 9 hours ago
This is clickbait speak. Stop talking like this or titling things like this. Dump what you are trying to say immediately, and put some flavor in it. Example
“LLMs.txt and MCP are useless pieces of shit, bad ideas.”
Forgeties79 | 9 hours ago
Statement: basic wordy elaboration.
Statement: basic wordy elaboration.
If this was actually a person writing, which I don’t believe it was, then they are using way too much AI and are actually being negatively impacted. They’re mimicking the tool that mimics us.
StilesCrisis | 7 hours ago
SomewhatLikely | 9 hours ago
Forgeties79 | 9 hours ago
wccrawford | 8 hours ago
Forgeties79 | 7 hours ago
rcymerys | 9 hours ago
I've had multiple marketeers try to convince me that it's a must for a website, but the best "proof" they could give me were some LinkedIn posts from other marketing consultants. I've tried replicating a scenario where Claude/ChatGPT would look for this file, but without luck.
0x696C6961 | 9 hours ago
drbscl | 9 hours ago
We’ve moved well past that, and even when docs MCPs were the best approach, everyone just used context7 or similar.
vorticalbox | 9 hours ago
because documenting your api in swagger is too much work?
canadiantim | 9 hours ago
drbscl | 9 hours ago
Context7 still helps for saving tokens though (ironically that's also the point of llms.txt, which the author also decries)
canadiantim | 6 hours ago
zodwick | 9 hours ago
deadbabe | 9 hours ago
petercooper | 9 hours ago