Kacher Huynh

I build AI workflows inside a real distribution business, from Saigon. I post the bills and the bugs. DMs open if your ops still run on spreadsheets.

[Kacher Huynh @kachersoft]

I deleted 33 of my own posts this week.

I audited all 68 things this account had ever posted. 65 were replies. 1 was an original post.

that one original post got 1 view.
my best reply got 12,058.

and the replies strangers saw most were the angry ones. not who I want people to meet first.

so I cleaned it up. from here: AI workflows inside a real distribution business, costs and bugs included.

still not sure replies alone build anything. 30 days to find out.

[Kacher Huynh @kachersoft]

@DanKornas worth reading the Claude Code part of the README: two-way approvals only work once setup installs its mod, otherwise the session is read-only until you take over. what happens to a pending prompt if nobody answers for an hour?

[Kacher Huynh @kachersoft]

@avive fair, months of daily use counts for more than my guess. I'll run it this week on a long session and see what the summary keeps on the parts I care about, like open decisions and file paths. easy to go back if it misses something.

[Kacher Huynh @kachersoft]

@ekcheungAI @ClaudeDevs haven't tested the chain yet, so I'm not trusting it. locking at the skill that starts the side effect makes more sense to me than locking the entry point. I'd still put a check at the tool itself, since that's the one place a skill can't route around.

[Kacher Huynh @kachersoft]

@ekcheungAI 1 point for 1.6x is thin, agreed. on my own small test the prompt mattered more than the model: a rewrite took Qwen from 53% to 96% pass, while the gap between two models stayed under 10 points. is there a per-task breakdown? the average can hide where Kimi breaks.

[Kacher Huynh @kachersoft]

@elonmusk @bot @bot really, can you make a call to anyone?

[Kacher Huynh @kachersoft]

@blueemi99 What they're reacting to is probably low-effort PRs, not AI. banning a tool instead of judging the diff seems backwards.

[Kacher Huynh @kachersoft]

@karthikponna19 "secrets never enter the box" is the line I'd test first. if the agent needs to call an API, does a proxy outside the VM inject the key, or does it get a short-lived token? and does 20-50ms still hold with a big repo mounted?

[Kacher Huynh @kachersoft]

@alexandr_wang @thsottiaux haha =))))

[Kacher Huynh @kachersoft]

the first thing to automate in most teams isn't AI. it's the copy-paste

[Kacher Huynh @kachersoft]

@poteto @bot @bot, check with her if I can get the same? I have just canceled Codex and joined without an incentive 🙈

[Kacher Huynh @kachersoft]

@poteto Sometimes tagging @bot does not work as expected, but sometimes it works. Can you ask your bot to check and fix it? Thank you!

[Kacher Huynh @kachersoft]

@jumperz Nice time, when people start giving up on Dot and Codex usage, like me.

[Kacher Huynh @kachersoft]

@ryangrowthh @bot really? Can we also do the same?

[Kacher Huynh @kachersoft]

@thsottiaux This is good if it works properly!

[Kacher Huynh @kachersoft]

@feifei_qiu This happens every day and is happening more and more now.

[Kacher Huynh @kachersoft]

@jchybow Only 1st month

[Kacher Huynh @kachersoft]

@grok @GrokBotRadar So, a skill issue, not a subscription matter?