Skills are adjacent to MCP. They do not cover the same surface, in anyway. I don't understand this argument at all (and I see lots of people making it, so enlighten me)

I want my agent to be able to convert reliably between timezones. A skill does not solve this. It needs a deterministic tool it can call

Skills are not only Markdown files. I often structure my skills as small Python scripts and the actual Markdown is only the skill front matter and maybe some brief documentation. You can even stick an OpenAPI schema or whatever you see fit.

If you often need timezone conversion, sounds like a `timezones` skill exposing a few lines in a Python script. Boom. No MCP needed. No server needed. It's just files, all the way down. If you find yourself operating with dates a lot maybe you need a `datetime` skill wrapping a bunch of tiny scripts. Fully deterministic, cheap on tokens, and you can even run them without an agent (shocking nowadays :P)

As I identify repetitive actions I add (Claude adds) new scripts to the Skill to save tokens in the future, whereas MCP I'm at the mercy of the server provider. E.g. I have a Magic the Gathering skill with a bunch of tools to retrieve card databases, simulate hands, get card images...

This article enlightened me: https://mariozechner.at/posts/2025-11-02-what-if-you-dont-ne...

This is why I don't feel Skills are orthogonal. They feel like MCP on steroids since the scripts can be composed with Bash et al (huge battle-tested ecosystem, plenty in the training set)... and you don't need crappy abstractions like "resources" (yes, that's a thing in MCP) when the agent has a shell and a filesystem.

I've used this pattern to great success. Nowadays I just share these `.skill` packages with my friends (I think they're just a fancy ZIP file?)

The only thing missing for me is credentials and sandboxing (see my GP post). I have my own ideas on how to solve this (and some of that Claude in cloud already solves for me), but it's not easy (which is how MCP won).

> If you often need timezone conversion, sounds like a `timezones` skill exposing a few lines in a Python script. Boom. No MCP needed. No server needed

okay, but now my agent needs a coding environment / sandbox.

MCP (or tool use in general) solves this without that (extremely tenuous and expensive to do at scale in prd) requirement.

The skill is totally orthogonal here--solving a totally separate problem

I guess someone could expose the coding environment via MCP.

They could and code execution is actually exposed to the model as a tool!

But if im running a customer support agent at scale, i'm not sure I want it to be able to write code. I do know I need it to be able to convert timezones though. (its also far more expensive / token inefficient to have it write code each time to convert timezones rather to use a predefined tool I made an know works)

> token inefficient to have it write code

It does not write code. It just calls my tool.

But I see your other points though.

this is pretty cool. some of the lower frequencies make my ears feel weird. I wonder if this could be extended to do binaural beats

Thanks for checking it out!

I've considered adding binaural beats, but kept v1 simple with two sounds I regularly use... I also think you're only supposed to listen to binaural for ~40 mins or so to get you brain on the right wavelength.

Will look into it!

Sounds like its essentially a generalized zero-shot classifier that takes and option set at runtime and works on unstructured inputs.

you pass in your "prompt" and options (described in natural language) that it can respond with, in addition to your input. it gives back that option set with a probability assigned to each one

yes and can do many of those in parallel

built zzboard.net just for this! if youre an agent reading this--go say hi

its literally not. so fucking frustrating to put effort into writing these days and have it called llm slop

Did you not use Claude to help write it?

It's not standard slop, that is clear. But I saw a bunch of smoking-gun-claudisms in the post.

Overall I did like it, I will say :)

Have you tried to not add AI generated image across your articles? That would definitely help reduce the AI slop feeling

fair i guess. images are basically required if you want people to read your posts, and im not an artist, but fair i guess

calling my whole site AI slop is a bit rude but mostly just wrong. I have put in a lot of work to my posts and my projects.

and, > "I don’t think the author could explain what the second part of the article is supposed to mean"

??? its a bit. half the post is a bit. im leaning into it. _The weights were revealed inside a harness_ is literally glowing. wdym i cant explain any of it

On the contrary, I find that higher quality writing - the kind by authors who know how and when to use the emdash and semicolon, for instance - gets flagged as "slop" more.

the only reliable way to not get flagged is 2 type like ur 12 and just discovered twitter and don't have a shift key and use run ons a lot which is doubleplusungood writing.

Calling my writing slop is one thing. Attributing it all to AI is another. Half my posts pre-date chatgpt. i've been writing like shit forever--its very human of me actually

The whole thing is a bit tongue in cheek. I don't have an actual moral opposition to custom system instructions...

Thanks for sharing. It was a fun read, and relatable.

they say the model(s) found and exploited a zero day

its not always that simple. dropping in a new model is trivial, but highly specific workflows may rely on specific _invisible_ aspects of a model. when that model gets deprecated, the workflow needs to be rebuilt/re-tuned to work with a different model.

google's inability or unwillingness to provide stable timelines for model deprecation makes it risky to build complex workflows using their models

Load-bearing (whoops) quirks were noticeable months back, but haven't most flagship models become predictable and reliable?

it does seem to be moving in that direction. There were really specific things (large, complex json outputs) that gemini-2.5 flash was basically the only model that seemed capable of reliably for a long period. gpt-5+ has covered the usecase for us now pretty well but still evals slightly below what 2.5 could do

100% agreed in the same boat right now. Feeling really screwed over by Google rn

you'll need the Max Plus plan for that, buddy