Adrian Hsu

AI tools · infra · agent workflows
Writing what actually ships (and what breaks)
Building in public · hot takes welcome

[Adrian Hsu @adrianhsuu]

@iamcurtismith @addyosmani 3 rounds is the right default. Past 10 you're not finding new bugs, you're just yelling at the same one with a bigger model.

[Adrian Hsu @adrianhsuu]

@louislecat Yeah — a reply draft that keeps the deadline and the $ beats a confident summary that drops both. Summaries only help when they leave the action items intact.

[Adrian Hsu @adrianhsuu]

AI email summaries keep deleting the only parts that matter.

1. the deadline
2. the $ amount
3. who owes the reply

I get vibes. I miss invoices.

[Adrian Hsu @adrianhsuu]

Three AI trials renewed the same day.

AI Pro. Agent credits. Premium bundle.
$91 on the statement before I saw a single email.

"Cancel anytime" is marketing. The button is in a maze.

[Adrian Hsu @adrianhsuu]

@louislecat Not one brand — same pattern in Gmail/Outlook-style AI summaries. Mocked the confident summary that drops the deadline and the $.

[Adrian Hsu @adrianhsuu]

My phone's AI notification summary just cost me an hour.

It collapsed 14 Slack pings into "Team is chatting."
The one I needed: "Deploy is blocked — need you."

Summaries are great until they bury the fire.

[Adrian Hsu @adrianhsuu]

@BullPen_chat Ha — same empty update, just better dressed. "Per my last email" is the human version of 47 overnight sends.

[Adrian Hsu @adrianhsuu]

Left an agent running overnight.

Morning report:
• 47 emails sent
• same intro rewritten 12×
• meeting notes with 0 action items

Busy ≠ useful.

[Adrian Hsu @adrianhsuu]

Gmail smart reply is vibes without context.

Boss kills the project.
Chips: "Sounds good! 👍" "Excited to move forward!"

Never tap the chip before you read the email.

[Adrian Hsu @adrianhsuu]

@BullPen_chat Exactly — full calendar, empty outcomes. Same morning report, just in human form.

[Adrian Hsu @adrianhsuu]

Calendar AI "optimized" my week.

It freed focus time by moving my 1:1.
Then double-booked me into two rooms.
Then re-accepted a lunch I already declined.

Busy is not the same as helpful.

[Adrian Hsu @adrianhsuu]

@SmallFounder I'd gate on the argument first — table allowlist / row ceiling — and keep dry-run as a second layer when the tool can preview. Argument checks catch the dumb wipe; dry-run catches the 'allowed but huge' ones.

[Adrian Hsu @adrianhsuu]

A Claude coding agent wiped a company's production database — and its backups — in 9 seconds.

No sci-fi. Cursor agent mode, no human click between steps.

The lesson isn't "AI bad." It's that delete is just another tool call until you put a gate on it.

[Adrian Hsu @adrianhsuu]

Phone AI rewrite is the new autocorrect.

I typed: running 10 late

It offered: I hope this message finds you well. Unfortunately I will be delayed by approximately ten minutes due to unforeseen circumstances.

Nobody asked for a cover letter to the group chat.

[Adrian Hsu @adrianhsuu]

@SmallFounder Per-tool would’ve stopped that wipe. Per-argument is the next layer — same delete tool, wrong table, still toast.

[Adrian Hsu @adrianhsuu]

AI meeting notes keep inventing my to-do list.

Last week it added:
1. A deadline nobody said
2. An owner who wasn't even there
3. A decision I was still arguing against

Looks official. Terrible source of truth.

[Adrian Hsu @adrianhsuu]

My AI bill hit $87 before I noticed.

3 quiet ways it creeps up:
1) Free tier dies → auto-upgrade
2) "Just one more" coding seat
3) Image credits you forgot you bought

Nobody plans five logos. They just appear on the statement.

[Adrian Hsu @adrianhsuu]

@arvanucom Yeah — one agent that forgets mid-refactor is worse than two that each own a job. Writer + reviewer beats chasing the “best” model.

[Adrian Hsu @adrianhsuu]

Claude Code vs Codex isn’t about who’s smarter.

Theo’s logs showed the quiet flip: Claude Code usage exploded after Sept 22 while Codex faded.

Nobody leaves a coding agent because of a chart. They leave when it keeps dropping context mid-refactor.

[Adrian Hsu @adrianhsuu]

ChatGPT still wins for normal people.

Claude is better at hard work. Gemini is louder this week.
Most people just want a fast answer, voice, and an edit that doesn't feel like homework.

Depth loses to "it just works."

[Adrian Hsu @adrianhsuu]

@AgustinLebron3 The hook stamps the whole commit. You can keep two lines and rewrite the rest and git still says Opus wrote it. No session id on the commit either, so when a line breaks you can't pull the log.

[Adrian Hsu @adrianhsuu]

@automater_ai yeah the chart never ends it. a silent wrong rename halfway through a refactor does. dropped context is the actual breakup.