Jonathan Hefner

Working on | All my issues are skill issues

[Jonathan Hefner @hefnerdotpro]

@andrewqu This one's for you:

[Jonathan Hefner @hefnerdotpro]

@evilrabbit_

[Jonathan Hefner @hefnerdotpro]

@andrewqu @RhysSullivan Not long enough.

[Jonathan Hefner @hefnerdotpro]

@rauchg I can hear the Sound Blaster 16.

[Jonathan Hefner @hefnerdotpro]

@andrewqu "Verchell". It's Italian.

[Jonathan Hefner @hefnerdotpro]

@badlogicgames @xovemnormie It is not required. Clients may choose to do it for a better UX. It is mentioned here: https://t.co/WMk1gdxL4H

[Jonathan Hefner @hefnerdotpro]

@badlogicgames @xovemnormie It's the HTML strict parsing debate from the late 90s, but for skills. Which of those is better UX for skill *users* (who might not be skill authors)? What is the cost of accepting malformed frontmatter (e.g., are edge cases tightly scoped)?

[Jonathan Hefner @hefnerdotpro]

@badlogicgames @xovemnormie The spec does say "YAML frontmatter followed by Markdown content". Skills authors should follow that. However, if a harness encounters malformed frontmatter matter, what should it do? What would provide the best user experience?

[Jonathan Hefner @hefnerdotpro]

@SomeCynic @badlogicgames See: https://t.co/VbhRfFg5J7

[Jonathan Hefner @hefnerdotpro]

@zeeg The honest answer is that it was copied from Claude Code Skills early on in the standardization process. Since then, CC has introduced even more features, which we decided to wait to standardize until they get more consensus. allowed-tools is a wart, but mostly inconsequential.

[Jonathan Hefner @hefnerdotpro]

@sarahwooders You're absolutely right to push back on that.

[Jonathan Hefner @hefnerdotpro]

@thdxr @kentcdodds

[Jonathan Hefner @hefnerdotpro]

@steipete

[Jonathan Hefner @hefnerdotpro]

@colmtuite samemaxxing

[Jonathan Hefner @hefnerdotpro]

@emollick I feel like that's at odds with Google's ethos. One: they aren't very product-centric, and two: they strongly embrace "the cloud is the computer". Seems more likely that their "everything app" will be Chrome.

[Elon Musk @elonmusk]

Grok now has skills

[Jonathan Hefner @hefnerdotpro]

@raw_works Very cool!! This is probably a dumb question, but why use Python for RLM? Why not Bash + (virtual) file system? To me, it seems like many models are currently trained to do RLM-like things in a Bash environment.

[Jonathan Hefner @hefnerdotpro]

@emollick In your opinion, is there a short list of places where people "get AI" more so than anywhere else?

[Jonathan Hefner @hefnerdotpro]

@sarahwooders I like it for tab completion in some cases. Typing `/com<TAB>` is clearly and obviously superior to typing `commit`.

[Jonathan Hefner @hefnerdotpro]

@mattpocockuk I've had this thought too. When the model is well familiar with a named concept, invoking the name nudges it toward that latent space -- very high leverage.

It's startlingly similar to mythology wherein knowing something's true name gives power over the thing itself.

[Jonathan Hefner @hefnerdotpro]

@evilrabbit_ Tangram