aaron@apirateslifefor.me

all things vibecoding

terms you should know

definitions

lint

/lɪnt/
definition

a tool, or the act of using it, that scans source code for likely bugs, style problems, and risky patterns without running the program. the tool itself is called a linter.

plain

an automatic proofreader for your code. it flags mistakes and sloppy spots before the app ever runs, so problems get caught early. people also say "a lint" to mean one such automated check.

in context

"i added a lint step to the build." → it wired in an automatic code checker that runs on every change, so bad code gets flagged before it ships.

ga'd

/ˌdʒiːˈeɪd/
definition

ga stands for general availability. to ga something, past tense "ga'd", is to release it from beta or preview into a stable, fully supported version available to everyone.

plain

it went from "try it at your own risk" to officially released and safe to rely on. if a feature is ga'd, it is finished and open to all users, not an experiment.

in context

"anthropic ga'd that in april." → it stopped being a limited preview in april and became the official, supported version anyone can use.

concepts

negative prompting / anti-negative prompting

definition

negative prompting is telling a model what to avoid or exclude rather than what to do, for example "do not mention pricing", or in image tools, listing what you do not want in the picture. anti-negative prompting is the counter-practice: because models tend to fixate on whatever you name, even to forbid it, you instead phrase everything as what you do want.

plain

telling the ai "don't do x" can backfire, because naming x can make it show up more. the fix is to say what you do want instead of what you don't. "write it in plain language" works better than "don't be wordy".

in context

"try rephrasing that as a positive instruction." → your "don't" rule is probably backfiring, so tell it what to do instead of what to avoid.

context poisoning

definition

when false, malicious, or low-quality information gets into a model's context window, the text it is currently working from, and then skews everything it produces afterward, because the model treats whatever is in context as true and keeps building on it.

plain

once bad info gets into the ai's working memory for a session, it keeps trusting and repeating it, so later answers get contaminated too. the usual fix is to start a fresh chat.

in context

"this thread looks context-poisoned, let's start fresh." → something wrong got into its memory for this conversation and is skewing the answers, so a clean session clears it.