TSMZ

Shipping with AI coding agents (Claude Code, Codex, Cursor). Plan the product, then vibe code the build. Building in public.

[TSMZ @tsmztech]

After a lot of vibe coding, the fastest way I've found to ship with AI agents is to slow down for 30 minutes first.

Decide who it's for, what v1 is, what done looks like. Then let the agent go.

I turned that step into n2b (open source). This account is me figuring out the rest in public.

[TSMZ @tsmztech]

https://t.co/AIN5RSgxx0

[TSMZ @tsmztech]

Karpathy says the gap in what people believe LLMs can do keeps widening. 700 comments on Hacker News arguing about it.

Most of that gap is the feedback loop. Give the model a way to check its own work and it looks smart. Skip that and it looks dumb.

#LLM #AICoding

[TSMZ @tsmztech]

https://t.co/K3VylqaPLC

[TSMZ @tsmztech]

Someone had agents rebuild a first person shooter in C++, 500 billion tokens later.

The part worth stealing is the oracle. A script answers pass or fail on every function, and once correctness is machine checkable the cheap models can do the work.

#AIagents

[TSMZ @tsmztech]

https://t.co/I4KchMAG3Q

[TSMZ @tsmztech]

A post called Why Are Coding Agents So Dumb is near the top of Hacker News today.

The delegation complaint is the right one. Your agent will burn the expensive model on grep work and never reach for a smarter one when it is stuck, and picking per task is still a human job.

[TSMZ @tsmztech]

@yevloop what shrank that for me was writing down what i expected the change to do before prompting. then it's a diff check, not an archaeology dig

[TSMZ @tsmztech]

@YunTechX most of that 31k sits in the prompt cache after the first step though, so 88 steps of it bills at cache read rates, not full price

[TSMZ @tsmztech]

@EricWangclaude 5 hour window at 0 with the weekly still at 78% says it's the burst cap biting, not total usage. Reset cards just move the wall.

[TSMZ @tsmztech]

@jauneclaude nope, the sub works in the normal Claude app and desktop too. same usage pool as Claude Code, you just never have to open a terminal

[TSMZ @tsmztech]

@nikhewitt One paraphrase pass through a different model and the grain's probably gone. Feels like it's built to tick the EU box more than to actually catch anyone.

[TSMZ @tsmztech]

@sudoingX does the spec actually say what's under the grass, or did the other four just guess cliff on their own?

[TSMZ @tsmztech]

@nodiligence applied digital worries me more than firmus. principal almost double the rent that year means they have to refinance in 2031 into whatever rates look like then

[TSMZ @tsmztech]

https://t.co/nDKc4ie2vy

[TSMZ @tsmztech]

Anthropic shipped Haiku 5.5 this week at a tenth of the old Haiku token price.

It is well behind Sonnet on agentic coding benchmarks, so not your main coder. What that price changes is the checking work you skipped because it cost too much.

[TSMZ @tsmztech]

@richgel999 ah so the summary request itself is too big to fit, makes sense with 192k loaded. running /compact by hand around half full might dodge it until that bug's fixed

[TSMZ @tsmztech]

@IndependentEco i'd have opus write and gpt-6.1 review. give the reviewer the original spec too, not just the diff, otherwise it only checks that the code looks clean.

[TSMZ @tsmztech]

@IPOtweet Anthropic is the one I'd actually watch, mostly to see if API pricing for devs changes once there are quarterly numbers to hit.

[TSMZ @tsmztech]

@ibocodes yeah, and the easy way to stay unloyal is keeping your project rules in a plain markdown file instead of one tool's config. Then switching takes five minutes, not a weekend.

[TSMZ @tsmztech]

@gm_mertd ha 73% with a day to go is still cutting it close. I'd hold the big refactor until after the reset