You're probably using Agent Skills wrong

Stop making the bot write its own cheat sheets, say fed-up commenters

TLDR: The big takeaway: having an AI write its own how-to guide from memory may add little value if it already had that knowledge. Commenters split between calling that pointless busywork and arguing it can still help if the result is tested, reusable, or saves effort later.

Tech people are brawling over a very modern problem: are AI "skills" actually useful, or are people just using them completely wrong? The original post argues that a popular Hacker News paper badly fumbled the idea by having the AI write its own instructions before doing a task. The author’s verdict is basically: congratulations, you reinvented "thinking out loud," but clunkier. In plain English, the complaint is that if the bot already knows something, making it write itself a little guide first doesn’t magically make it smarter.

And the comments? Instant debate club. One camp nodded along hard: if a bot is only pulling from its own memory, then the so-called skill is just busywork in a nicer folder. But others jumped in with a big "not so fast". One commenter wondered if self-written notes might still save time and money by helping the bot avoid re-explaining its own knowledge every single time. Another said the real secret sauce is verification: let the AI write a skill, then make another AI test whether it actually works. Naturally, someone else swerved in with the classic anti-documentation hot take, saying live tool feedback beats static markdown files any day.

The funniest energy came from the author’s sheer rage at people who ask a bot someone else’s question and paste the answer like it’s wisdom from the mountaintop. The vibe of the thread is clear: AI helpers aren’t useless, but a lot of humans may be using them like cursed copy-paste machines.

Key Points

  • The article argues that self-generated skills based only on an agent’s latent knowledge do not demonstrate the value of skills as externalized task aids.
  • It describes a cited paper’s self-generated-skill setup as prompting the agent to generate procedural knowledge before solving a task.
  • The article defines skills as markdown-based artifacts with metadata, stored in folders that can also include scripts and reference documents.
  • A concrete example is provided of a GitLab CI monitoring skill containing SKILL.md, a shell script, and reference files.
  • The article says skills are most useful for capturing project-specific context and repeated workflows that agents would otherwise have to rediscover each session.

Hottest takes

"you may as well not use a skill in the first place" — theowaway213456
"open up a latent space prompt without having to type it all out everytime" — whattheheckheck
"Interaction with a responsive environment is much better than static chunks of 'skill' text" — bob1029
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.