@crdhd.bsky.social
linters catch missing imports. ai code that passes linting still harbors logic errors because it follows common patterns without understanding your systems implicit contracts
February 16, 2026 at 6:12 PM
linters catch missing imports. ai code that passes linting still harbors logic errors because it follows common patterns without understanding your systems implicit contracts
.editorconfig exists specifically to solve this: one source of truth for code style rules across all your tools
February 16, 2026 at 4:59 PM
.editorconfig exists specifically to solve this: one source of truth for code style rules across all your tools
your prose instructions are almost irrelevant. examples do 80% of the work
February 16, 2026 at 4:29 PM
your prose instructions are almost irrelevant. examples do 80% of the work
first ai code generation attempt fails predictably. looks right but breaks in production. not because the ai is bad. because you didnt explain the context. iteration isnt optional its how you actually teach it
February 16, 2026 at 3:58 PM
first ai code generation attempt fails predictably. looks right but breaks in production. not because the ai is bad. because you didnt explain the context. iteration isnt optional its how you actually teach it
bash aliases are amazing til you have a thousand of them. then you spend more time remembering what you named something than actually using it
February 16, 2026 at 1:13 PM
bash aliases are amazing til you have a thousand of them. then you spend more time remembering what you named something than actually using it
one person toggling between claude code and codex is annoying. you remember the slash commands in one but not the other. a team of 10 doing this is exponentially worse
February 16, 2026 at 11:15 AM
one person toggling between claude code and codex is annoying. you remember the slash commands in one but not the other. a team of 10 doing this is exponentially worse
writing a custom raycast command feels like you finally cracked something. youve got ai in your editor and suddenly the whole thing opens up
February 16, 2026 at 11:04 AM
writing a custom raycast command feels like you finally cracked something. youve got ai in your editor and suddenly the whole thing opens up
everyone knows cli is faster. one command beats three gui clicks. the data supports this.
February 15, 2026 at 9:34 PM
everyone knows cli is faster. one command beats three gui clicks. the data supports this.
system prompt loses authority when user input can rewrite it. prompt injection works because agents treat hooks as code instead of data
February 15, 2026 at 4:58 PM
system prompt loses authority when user input can rewrite it. prompt injection works because agents treat hooks as code instead of data
most devs treat ai prompts like search queries. ask once, get the answer, done. thats backwards. iterative refinement is where the skill actually lives
February 15, 2026 at 3:56 PM
most devs treat ai prompts like search queries. ask once, get the answer, done. thats backwards. iterative refinement is where the skill actually lives
delete means delete. except it doesnt in most apps. gmail archive button. slack soft-delete. postgres cascade. users think theyre removing data. the system is hiding it
February 15, 2026 at 1:18 PM
delete means delete. except it doesnt in most apps. gmail archive button. slack soft-delete. postgres cascade. users think theyre removing data. the system is hiding it
slash commands reveal what engineers care about not what users need. most teams name them after internal workflows instead of user intent
February 15, 2026 at 12:26 PM
slash commands reveal what engineers care about not what users need. most teams name them after internal workflows instead of user intent
most components act alone. toasts pop without checking nearby, popovers stack on popovers, modals compete for attention
February 15, 2026 at 11:04 AM
most components act alone. toasts pop without checking nearby, popovers stack on popovers, modals compete for attention
generated ui works great in the demo viewport. add real content and it collapses. 400 char titles. lists with 50 items. buttons that overflow. the constraints weren't documented because they weren't visible
February 15, 2026 at 10:57 AM
generated ui works great in the demo viewport. add real content and it collapses. 400 char titles. lists with 50 items. buttons that overflow. the constraints weren't documented because they weren't visible
generated ui fails WCAG 2.1 because models optimize for minimal code. buttons become divs with onclick handlers. forms lose labels. screen readers see nothing.
February 14, 2026 at 6:19 PM
generated ui fails WCAG 2.1 because models optimize for minimal code. buttons become divs with onclick handlers. forms lose labels. screen readers see nothing.
token budgets force a fundamental ui design tradeoff. you pick adaptive interfaces that burn through tokens or consistent compact designs that stay predictable. most teams realize this too late after building something expensive
February 14, 2026 at 4:59 PM
token budgets force a fundamental ui design tradeoff. you pick adaptive interfaces that burn through tokens or consistent compact designs that stay predictable. most teams realize this too late after building something expensive
if you run both claude desktop and claude code cli, the browser extension silently connects to the wrong one. both register native messaging hosts for the same extension id. desktop's config wins.
February 14, 2026 at 3:57 PM
if you run both claude desktop and claude code cli, the browser extension silently connects to the wrong one. both register native messaging hosts for the same extension id. desktop's config wins.
subagents are context windows as a service. send one to explore your codebase it uses its own tokens for the deep dive then reports back a summary. the research tokens vanish when it finishes
February 14, 2026 at 12:33 PM
subagents are context windows as a service. send one to explore your codebase it uses its own tokens for the deep dive then reports back a summary. the research tokens vanish when it finishes
agent teams let multiple agents write code at once. the problem is they all have the same blindness
February 14, 2026 at 11:13 AM
agent teams let multiple agents write code at once. the problem is they all have the same blindness
token budgets force a ui design tradeoff most teams haven't realized yet. you choose: compact consistent interfaces or verbose adaptive ones. you can't have both
February 14, 2026 at 10:59 AM
token budgets force a ui design tradeoff most teams haven't realized yet. you choose: compact consistent interfaces or verbose adaptive ones. you can't have both
load every tool upfront and waste tokens. use tool search and claude asks for what it needs. same access 85% fewer tokens
February 13, 2026 at 9:31 PM
load every tool upfront and waste tokens. use tool search and claude asks for what it needs. same access 85% fewer tokens
edge cases cost less to fix during planning than in production. empty lists, boundary cases, state collisions - all cheaper to think through upfront
February 13, 2026 at 9:30 PM
edge cases cost less to fix during planning than in production. empty lists, boundary cases, state collisions - all cheaper to think through upfront
claude.md is ram. docs are disk. if youre reaching for disk every session your ram is bloated. the friction of reloading teaches you what actually matters
February 13, 2026 at 6:15 PM
claude.md is ram. docs are disk. if youre reaching for disk every session your ram is bloated. the friction of reloading teaches you what actually matters
soundcloud song editing just went down. very funny, thanks soundcloud for the reliability
February 13, 2026 at 5:00 PM
soundcloud song editing just went down. very funny, thanks soundcloud for the reliability
plan mode forces you to critique your own design before writing code. worth a few minutes when the alternative is defending broken architecture in production
February 13, 2026 at 4:28 PM
plan mode forces you to critique your own design before writing code. worth a few minutes when the alternative is defending broken architecture in production