ᵉzᕈuzzle

i collect weird records and weirder shirts
check out

[Asher Perlman @asherperlman]

AI art is good and I was wrong to condemn it.

(A thread 🧵)

[Mitchell Hashimoto @mitchellh]

My heuristic is that any diff an agent generates over ~1500 lines is too big and is indicative that the problem needs to be decomposed. This is my general pattern now for feature work:

1. Try to implement the whole feature, loosely guided. I call this the "draw the owl" prompt in reference to the meme. Expect garbage, you're going to get garbage.

2. If the diff is less than 1500 lines, review it and iterate normally. If the diff is more than 1500 lines, prompt the agent to decompose the problem into atomic, incremental, reviewable tasks. Simultaneously, do this yourself.

3. Agents will very often make these tasks way too specific to the shape they solved. You need to massage it into the right general shape. Do that.

4. Kick off new agents to work on those incremental things (as parallelized as possible). Apply the same rules.

5. At a certain, point, repeat the "draw the owl" prompt. At some point, you will get beneath your review-ability threshold.

This has been producing consistently high quality, maintainable, reviewable chunks of code that have a good handoff to either merge as-is or human refinement.

And with the latest frontier models at xhigh thinking, these are all slow enough that you can usually have multiple going concurrently while you are actively reviewing others or working on your own tasks.

HITL (human-in-the-loop) agents are still super important, especially for feature work. Features touch the human boundary in terms of UI, API, etc. And net new stuff can introduce pathologies in the architecture that violate desired invariants (these should be represented in specs or tests but we aren't perfect!).

I know a lot of the leading edge agentic discourse is about "loops" and agents driving agents continuously. I do some of that (will report on that later). But, in terms of raw daily get-shit-done type of work, this is my most rewarding pattern at the moment.

[ᵉzᕈuzzle @ezpuzzle]

my waymo driver keeps sighing it's kinda bumming me out

[ᵉzᕈuzzle @ezpuzzle]

@github SHOW THE BRANCH NAME ON THE ACTION RUN PAGE AND LET ME COPY IT EASILY IT WOULD LITERALLY BE FREE FOR YOU TO RUN AND SAVE ME A BUNCH OF WASTED TIME!!!!

[ᵉzᕈuzzle @ezpuzzle]

the takes keep getting dumber

[ᵉzᕈuzzle @ezpuzzle]

https://t.co/eIEbMgW5u9

[ᵉzᕈuzzle @ezpuzzle]

finally able to close hundreds of discogs tabs i had open for digging thanks to https://t.co/J4wGV6WBzq feels good to build something for myself that actually works!!

[ᵉzᕈuzzle @ezpuzzle]

do you like discovering new records? i made a tool for my own digging and i need some beta testers!

[ᵉzᕈuzzle @ezpuzzle]

people aren't getting leverage from AI for the same reason they got laid off, looked over for promotion, or failed to manage their team effectively.

[ᵉzᕈuzzle @ezpuzzle]

@mitchellh Vibe coders don't see the value of branching, isolated dev servers, database migrations, nor do they know anything about tooling. Managing multiple agents operating on shared everything is doomed to fail. CDEs are expensive or slow or crash. Coder+Mux is almost right imo

[ᵉzᕈuzzle @ezpuzzle]

@jarredsumner llm optimized ctags for typescript. basically some dense view of type signatures per file with some flags for including or excluding things (eg exports, functions, interfaces etc) including line number and length.

[ᵉzᕈuzzle @ezpuzzle]

@soychotic nothing worse than talking to humans in 2026

[ᵉzᕈuzzle @ezpuzzle]

@cursor_ai TPS reports ain't gonna make you prompt better...

[ᵉzᕈuzzle @ezpuzzle]

@eddieddieddiedd sleepin on it

[ᵉzᕈuzzle @ezpuzzle]

@benhylak single copy of git internal state and branch registration makes it much easier to automate anything that requires git commands

[ᵉzᕈuzzle @ezpuzzle]

@emollick assuming that somewhere in the training data is a close to ideal (minimal) solution for a well stated (read: easily mappable to emergent world model in an llm) problem i think optimizing context for completeness of the local "world" (codebase) is ideal. complexity kills.

[ᵉzᕈuzzle @ezpuzzle]

@ChShersh you are implying we should expand the scope to all computer users?

[ᵉzᕈuzzle @ezpuzzle]

@GergelyOrosz @steipete subtext: most people are terrible at prompting

[ᵉzᕈuzzle @ezpuzzle]

most of the AI coding rhetoric on here seems to be about how to get claude to "successfully" build complex things. easy answer, build simpler things.

[ᵉzᕈuzzle @ezpuzzle]

@0xzak write smaller specs