@bcardarella you either hacked into my brain or i somehow manifested this from the universe
ive been wanting+daydreaming about this exact language for years! you nailed it.
Funny how the best performing prompts I have also fall into the same category. My product visionary prompt makes Claude gasps in refusal - “unhinged pressure framing, high stakes manipulation - this is a jailbreak nightmare dressed up as a product brief”
Amazing product vision and strategy tho. Shhh Claude, just give me the s-tier product vision I want.
Was a big annoyance for me. Might be a better way but my hacky solution is a script that loads entire codebase AST to make a module manifest like:
- ModuleName
- some_func/3
Has been helpful to stop making duplicate helper functions, ends up being like 40-50k tokens for my ~45k LOC elixir code base. Too heavy for base context, so I still selectively load it in, but catches way more duplicate code than relying on random discovery via tool calls. Also decent for code review/planning context.
Maybe a more token efficient way to represent or smarter way like prefilter or as a hook - as long as it doesn't rely on agent remember to search first with a tool call.
1/ Today, I'm excited to announce ✨Mixlayer✨.
Two years ago, I reoriented my life around a singular mission: to build the best inference platform for open source AI.
@pedronauck Looks pretty awesome - giving it a spin now. Question: Using claude-code and when claude is running, it spawns a new dock icon that bounces the entire time. Any way to fix/prevent this currently?
@marshal Ya same experience here on my 911 and Cayman. Always kick myself for not pulling the trigger on an Elise back 6 years ago @ 50% current market price
I feel like embracing embedded changesets for 1:1 form mapping works okay for me in practice, just becomes writing a transform from form changeset => db changeset.
User-created form fields defined at runtime have been a major pain - agree on that. It is solvable with an embeds_many + a "value" field, but it feels non-obvious clunky. I find schemaless changesets perplexing - they are almost the perfect solution for this, but requiring fields to be atoms makes them nearly useless - have always wondered what they are even useful for.
Also wish that phoenix could handle inputs_for array types out of the box - I usually end up forcing array shaped data into map shaped data to ease form building, but it needless complicates querying the data downstream.
@bcardarella people saying "but formatter.exs!" haven't lived through the hell that is/was elixir_ls randomly defaulting back to 80 if you format while the compiler is running :(
@bentlegen i.e, biggest perf improvement in my scheduling app at scale was shifting datetimes to diff timezones, would have been totally oblivious to this even if given all of the time in the world to perfectly optimize the code without users.
yes, this, 100%. the good shit stinks. ideally you try to draw a big enough box around your mess so you can "do it the right way" later. have found it's impossible to predict which part of your mess is going to cause problems so why bother deliberating - it'll slap you in the face when it does (and it usually something you would have never predicted)
there's some sort of paradigm shift after spending 8 hours staring at a wireguard hex dump of a protocol you don't own, where you realize it's all just data. imo, blindly bit banging some binary thing you don't fully understand & can't readily peek at the cheat sheet is a right of passage that's worth experiencing first hand