crdhd.bsky.social
@crdhd.bsky.social
shift code review from does this compile to does this do what the system expects in context. thats where the real mistakes live. not in syntax but in implicit contracts
February 16, 2026 at 6:12 PM
plausible-but-wrong is the failure mode. code looks idiomatic so it passes review. but it encodes assumptions that are invisible until load hits or edge cases arrive
February 16, 2026 at 6:12 PM
functional correctness requires execution validation not just parsing. unit tests are how benchmarks like HumanEval verify code actually works. running it matters more than reading it
February 16, 2026 at 6:12 PM
use editorconfig for what its designed for. dotfiles and version control for the rest. one tool for each problem
February 16, 2026 at 4:59 PM
but style alone doesnt solve it. editorconfig handles formatting but not hooks commands or model selection. those still need separate config
February 16, 2026 at 4:59 PM
indent_size line_length trim_trailing_whitespace. define once in .editorconfig and both claude code and codex apply the same rules. no duplication no drift no wondering why the formatter disagrees with the linter
February 16, 2026 at 4:59 PM
spend your energy finding or writing the right examples. thats where the leverage is. prose polish is the wrong optimization
February 16, 2026 at 4:29 PM
when examples are solid your instruction text barely matters. the pattern is in what you showed not what you said
February 16, 2026 at 4:29 PM
each example in a prompt is called a shot. show a model 3-5 good ones and performance jumps. most devs write paragraphs expecting the ai to infer. thats backwards
February 16, 2026 at 4:29 PM
the best tool is the one you don't have to remember exists. discoverability won. and bash aliases lost not because they're slow but because they require remembering
February 16, 2026 at 1:13 PM
google made memorizing urls stupid. typing a url became an act of faith. you search instead. raycast did the same thing to your command toolkit
February 16, 2026 at 1:13 PM
raycast doesn't ask you to remember. 1500+ extensions in one place. you search, filter, find. same speed as aliases but zero memory tax
February 16, 2026 at 1:13 PM
the real cost isnt maintaining two tools. its losing the ability to enforce consistency and share learned patterns with your team. entropy through fragmentation not incompetence
February 16, 2026 at 11:15 AM
each team member now needs to remember which tool theyre using and what config it has. standardization becomes impossible. pattern-matching breaks because the patterns are different per tool
February 16, 2026 at 11:15 AM
claude code has .claude/commands, CLAUDE.md project memory, /.hooks. codex has setup-scripts, actions, config.toml. nothing overlaps. so you learn custom workflows in each tool independently
February 16, 2026 at 11:15 AM
the win isnt better reasoning. its asking better questions. mechanical work it handles. novel architecture is still yours. thats the design
February 16, 2026 at 11:04 AM
teams that won didnt try to be clever. they treated raycast like a ui over existing tools. custom prompts handle the 20%. everything else uses built-in
February 16, 2026 at 11:04 AM
then you ship it and realize youre just wrapping api calls. the actual logic still lives in your scripts. fancy curl with delusion
February 16, 2026 at 11:04 AM