Nick Lian

Using AI to make more room for being human.
AI Builder | DM for work & collabs
📧 : [email protected]

[Nick Lian @NickkkLianSYCC]

Seven major AI launches in three weeks.

Building with these models, I keep running into the same question: will the next release make what I just built pointless?

This is my attempt to name that feeling, and to work out what still keeps its value. https://t.co/wxxzwcVjNQ

[Nick Lian @NickkkLianSYCC]

Cory Doctorow’s new book on AI splits people into centaurs (the machine helps, you set the pace) and reverse centaurs (the machine sets the pace, you check its output).

Same tool, opposite jobs. The difference is who decides what gets checked, and how much.

[Nick Lian @NickkkLianSYCC]

@ThePrimeagen Five minutes is the worst possible wait: too long to stay focused, too short to start something else. Either it gets fast enough to stay in the conversation, or slow enough that you hand it off and come back to a finished answer.

[Nick Lian @NickkkLianSYCC]

@rauchg Makes sense: once the frontend is cheap enough to regenerate, all the care moves to the data and APIs underneath. Just-in-time UIs only stay trustworthy if the backend is boring, versioned and hard to break.

[Nick Lian @NickkkLianSYCC]

@claudeai The interesting part is where the line sits. Risky is easy to define for deletes and payments. It's harder for sending a message or submitting a form, where the risk isn't what the action breaks but who it speaks for.

[Nick Lian @NickkkLianSYCC]

https://t.co/wMy7COlgRi

[Nick Lian @NickkkLianSYCC]

the 1,011 print as 360 lines: a repeated value is listed once.

still red: the 28 home paths that name a real account and 66 key-shaped strings, both picked by hand. e-mails and phones weren't checked that way.

0.1.7, MIT: https://t.co/RzkEUoYQy1

a star helps if it's useful

[Nick Lian @NickkkLianSYCC]

5,560 red lines: what my privacy gate printed on 20 public agent-skill repositories.

mostly wrong. 4,135 of its 4,983 "e-mail addresses" were package@version strings. 298 of its 326 home-folder path alarms named nobody.

after the fix: 1,011, and it still fails 18 of the 20.

[Nick Lian @NickkkLianSYCC]

@pixel2ai Great, but “date” is a human construct 📅

[Nick Lian @NickkkLianSYCC]

@pixel2ai how do you prove we are going forward instead of going back ;)

[Nick Lian @NickkkLianSYCC]

7 rules, on by default. 5 came from things that went wrong here.

bash only. it doesn't stop git reset --hard or git clean; cc-safety-net does.

stdlib python, MIT: https://t.co/GAJXJvGZ0x

if it's useful, a star helps me know what to build next!

[Nick Lian @NickkkLianSYCC]

git add, commit, push: all three exit 0. the commit records every file in the repo as deleted.

so i built a claude code hook that stops before that push and 6 other commands, says which incident the rule came from, and asks me first.

the screenshot is that push on a test repo.

[Nick Lian @NickkkLianSYCC]

@omkarships Me

[Nick Lian @NickkkLianSYCC]

Did the math on Claude Haiku 5.5 pricing. Three things the launch posts skip:

1. The 100k line is a cliff, not a ramp.
A 99k-token prompt bills at $0.10 per million input tokens. At 101k, the whole prompt moves to $0.50. Same prompt, 5x the price. Output does the same: $0.50 → $2.50.

2. "10x cheaper than Haiku 4.5" is per token, not per job.
Haiku 5.5 uses the newer tokenizer, which counts about 30% more tokens for the same text (Anthropic's own number). So for the same text:
• under 100k: ~7.7x cheaper (about 87% less)
• over 100k: only ~35% less

3. The cheap tier is smaller than it looks.
100k tokens on the new tokenizer is roughly 77k tokens of Haiku 4.5-sized text.

It is still the best deal in the lineup: 20x cheaper than Sonnet 5.5 per token under 100k, 4x over.

What I'm taking from it for agent setups: lots of small, fresh sub-agents, each call well under 100k. One long-running helper that drifts past the line pays 5x for everything it reads.

Numbers are from Anthropic's pricing page and the Haiku 5.5 model page; the division is mine.

[Nick Lian @NickkkLianSYCC]

@cheaf25master In this transition stage, no. Few years later, yes, but with brand new ways to learn.

[Nick Lian @NickkkLianSYCC]

@BabaDev_ I only build things that I need and have to use.

[Nick Lian @NickkkLianSYCC]

@pixel2ai Funny how AI might be what pushes humans back toward being human.

[Nick Lian @NickkkLianSYCC]

@rileybrown The better the model, the more prompting becomes coaching: set the vision, create space, get out of the way.

[Nick Lian @NickkkLianSYCC]

@nikitabier Identification is only half the problem imo.
If “I’m an agent” = worse access, agents are economically incentivized to impersonate humans forever.
The real standard probably needs to cover identity, permissions, delegation, pricing and liability.