Pieter Levels says skills features in AI coding agents are overrated and equivalent to basic text files, preferring direct instructions on tasks and methods instead

@levelsio I did like the three.js skills for vibejam tho. They have so much knowledge i don't have!

@levelsio I recommend this article. It talks about the overuse of skills, their trade-offs, and when to actually use them so they provide real value.

https://efexen.substack.com/p/does-your-skill-earn-its-keep

@levelsio Skills are like functions.. If you need it once, prompt it. If you find yourself explaining it multiple times, create a skill

Not in the slightest. Actually a large part of the reason I appended that last sentence.

But it’s not that hard to grok the architecture of how LLMs work or the fundamentals of the SDLC

Giving text to something one way isn’t that fundamentally different to giving it text another way… which is all you’re doing here… but it’s pre written for easy repeatability that maintains flexibility for when to use it.

Same, but these are useful. RTK -> appends to each command and get rid of the unnecessary things from it automatically -> less token usage Caveman -> removes unnecessary blablabla from to output -> straight to the point Superpowers -> for complex features it is sometimes better in the planning phase

I love skills!

Most skills are just elaborate system prompts and a way to make & share custom LLm memory

Also way better than 99% of MCPs, just point it at an api and tell it to make a skill and now you have your own custom MCP

Helpful because I don’t want to type the same long commands over and over

Also nice because you can make it automatic, kinda like training your ai to do something again later

Becoming popular in big companies since you can have many people making skills and sharing them, now your agent can do things you didn’t even know existed

@levelsio skills are useful for workflows and progressive disclosure of knowledge model wasn't trained on, e.g. "make-hyperframes-video-in-my-brand-style".

but most skills like "marketer" or "growth-hacker" are beyound useless, they're harmful.

I didn’t either until I decided to abstract all my business metrics as an api and told a skill how to do it

Now I say “use Chrome to audit my Google Ads account, then figure out based on my company metrics using $companyskill if I’m optimized or if any changes need to be made, compare it to seasonality data”

Allows it to find recent COGS data, performance, etc and revise ads/road targets etc

Also do this by having it browse my Google Analytics accounts and find emerging keywords to implement in my website copy and/or launch SEO pages based on the intelligence it finds, but that’s not a skill

Anyway like you I don’t really use skills until I built this one

@levelsio Nice i use Context7 for that, they make docs LLM's can read more easily.

@levelsio Have you tried them: https://github.com/everyinc/compound-engineering-plugin

Had pretty great success with them and their dev logic

@levelsio The only skill I find invaluable is taste-skill

I describe my product idea and have it by generate 3 sample renders using $imagegen-frontend-web

Then have it use my favorite render of the options to build with $image-to-code

@levelsio Skills are just for when you want to bottle/save some way of doing something

@thijsmakes Ah yes that makes sense, sometimes I just copy paste documentation or a URL to docs

@levelsio I had the same question so I added evals for my french accounting/paperwork skills repo. It runs scrnarios multiple times with and without the skills and compares the results. Conclusion: yes skills are very useful.

@levelsio Skills are for people who are too inexperienced / stupid to tell their AI what to do. They revel in being able to “install a skill” to use a well-documented API / SDK that they wouldn’t even know exists.

If you need to use a skill, that’s a skill issue (99% of the time).

Skills work really well for capturing tone-of-voice. At Loop Earplugs we have to produce A LOT of copy.

If you show Claude a lot of examples and the ask it to encode the patterns and nuances in a Skill, we manage to get copy for all our marketplaces 90% right. Works even better than fine-tuning

@levelsio I have a skill that whips up a new devbox on hetzner with everything I need preconfigured, automatically enables all firewalls and adds it to my tailnet

all I need to do is tell codex or Claude “set up a new devbox” and it knows exactly what to do

@NevF @levelsio Nah, it's not you, it's mostly people that want to have a "magic button" in their setup

Check this tool to see which skills are not useful in your setup: https://github.com/av/skilled

@levelsio ive been analysing what they do a lot, often it looks up skills automatically and yea i should turn it off because its a verbose waste of tokens eg in this example when it looks up how to use git when it knows already how to use git its in the weights....

@levelsio I used to think the same.

But when for example, after repeatedly building endpoints on an existing backend, I started asking Codex:

“What have you learned from this workflow?” And then “Please turn it into a skill.”

That became an exponential improvement in my workflow.