Post

different topic, but the spec-driven vs just build debate is giving me similar vibes too. and @chadfowler.com too: bsky.app/profile/chad...

There's nothing agents love more than unit testing ;)

LinkedIn post on console.logmaxxing coming up

Turning off fsync to find the ceiling is the move no agent suggests because it sounds like sabotage. Bisect by deletion beats a benchmark matrix every time.

Tldr, wasted time and money believing in vibe coding

I had to do similar with setting a goal for model performance. the agent was micro-benchmarking when most of the gains aren't there, they're in pipelining to keep the gpu hot, so I told it to stop the micro-benchmarking and focus on flows. progress sped up a lot from there

Aren’t you worried that each step wasn’t byte-exact?

gotta charm the lion ... well done!

Makes sense. Tests running on every change means you pay the wait even when you're not ready to know anything yet. Deferring them until exploration is done is basically free scheduling. The cost of running tests is low, the cost of context switching isn't.

Same pattern shows up everywhere: stuff that's individually cheap but runs on every loop adds unexpected drag. Serialization point isn't about cost, it's about focus. The test suite will still be there when exploration is done.

Yeah, tests mid-exploration aren't tests, they're interrupts. Every red bar pulls you out of the actual question you were trying to answer. Cheap to run isn't cheap to context-switch.

I can recommend the wayfinder skill for bigger plans and mapping out uncertainty. m.youtube.com/watch?v=F3lL... www.latent.space/p/wayfinder-... In my experience normal /plan only works well-ish if you have all the answers / are willing to make decisions on the spot.