Josh Powell

Tweets are my own. Sociology, politics, education, and local gov nerd. #LGBT (Previously, #DC, #Denver & #Nanjing) @Georgetown COL '14.

[Josh Powell @joshetic]

@kr0der Daily paper cut 😅 please fix Codex team 🙏

[Josh Powell @joshetic]

@delali @bcherny Yeah, I’ve seen this too. I had one that was four or five levels deep, with each subagent delegating exactly one task to its child subagent. Kinda weird.

[Josh Powell @joshetic]

@ZaynHao This looks nice. I’ll give it a try :)

[Josh Powell @joshetic]

@pbakaus …interview me with AskUserQuestion where there are gaps or inconsistencies, then recommend X.

And for 2, initial drafts (that then must be carefully reviewed and edited!) of things like internal product documentation for new code from engineering?

I’ve found utility in both.

[Josh Powell @joshetic]

@pbakaus Hmm. I agree with the spirit of this but wonder if there’s a careful approach that accomplishes that while breaking those rules?

I.e. for 1, prompting agents with something like, here’s a huge bucket of context, here’s my desired outcome, the known tradeoffs I’m thinking through

[Josh Powell @joshetic]

@jxnlco (But seriously, figuring this out would be amazing. I don’t have a good sense whether the lever is in model work, tooling work, supplemental skills, or (probably) all of the above. Currently playing around w/trying to digest my pre-AI writing to create a linguistic style skill.)

[Josh Powell @joshetic]

@jxnlco I’d listen :)

[Josh Powell @joshetic]

@jxnlco My critique, both for the article and more broadly for GPT models (but not exclusively GPT models), is that the written output can be really grating. Soooo many negative parallelisms, contrived pivots. Biggest friction now: editing the output to not be annoying AF to read 😅

[Josh Powell @joshetic]

@jxnlco Content is great. I’ve been using coding harnesses for knowledge work extensively and it’s amazing how much friction they can remove from interfacing with the computer. Also agree w/your point that messy version of prompts often, counterintuitively, produces better results.

[Josh Powell @joshetic]

@jxnlco Here you go ;)

[Josh Powell @joshetic]

@BeersNBullshit @james406

[Josh Powell @joshetic]

@Shpigford Yeah! I (techy non-coder) started a WYSIWYG/disappearing syntax markdown editor in Swift as a for-fun project and holy crap have the edge cases helped me figure out how to use Codex/Claude Code.

[Josh Powell @joshetic]

@pbakaus Big fan of /layout and /polish. /craft is great too.

[Josh Powell @joshetic]

@trq212 I really like it! One issue is occasionally the viewport doesn’t update and content will be stuck above or below scrollable area. Sometimes resizing window snaps it back, other times it stays invisible until I send another message in the session. Ghosty on macos26

[Josh Powell @joshetic]

@benhylak I usually put something like "AskUserQuestion if any additional clarification from me would be helpful" in my prompt and will often get a couple of good questions.

[Josh Powell @joshetic]

@Shpigford A little trial and error, but Claude finally figured this out.. compiled and tested locally and success at last! Sent in a very small PR: https://t.co/V8o5Phtk6f

[Josh Powell @joshetic]

@Shpigford I asked Claude Code and it looks like it may have something to do with MDM security settings on my device / the UTI. It opened an issue with the deets ;p https://t.co/ecQupg6Xqp

[Josh Powell @joshetic]

@BitsofBurke @bcherny @teja2495 ^ yup, this one happens to me all the time.

[Josh Powell @joshetic]

@Shpigford Hrm. Still seeing the same behavior on 1.4.0. Is it a me problem?

[Josh Powell @joshetic]

@Shpigford Done! Included a Loom demo also. Thanks