Ben Carlsson

assigning an agent a task has made me appreciate how much context i normally communicate by pointing at a screen and saying “that”

the fastest way to learn which parts of your app are hardcoded is to make the window smaller

terminal colors are fun because your carefully chosen gray might be someone else’s carefully chosen background

every settings page contains at least one checkbox whose consequences can only be learned experimentally

what’s the smallest piece of evidence that makes you go “okay, the agent actually checked”?

a good default is a decision i never have to learn was available

i want the program to remember my preferences with the persistence it currently reserves for remembering the wrong window size

k

a coding agent reading the readme before doing anything is setting an uncomfortable precedent for the rest of us

a keyboard shortcut is not intuitive just because the letters are somewhere in the name of the feature

a dialog that opens under my cursor has made a decision on my behalf

i would like “resume” to include the part where i remember why i was doing this

i can tell a coding agent exactly what i want as soon as it shows me something i don’t want

which bit of your development setup would be hardest to explain to a person borrowing your laptop?

what’s something you still do manually because explaining it to an agent feels harder than doing it?

LLMs are like mad libs but every word is blank

i want fewer occasions to say “wait, before you do that”

what is your most unreasonable expectation of a text box. i am prepared to take it seriously

the computer can do the tedious part. i would like to stop being the courier between tedious parts

the software knows i have unsaved text. i would love for this knowledge to influence its decisions

somewhere a terminal is reporting this key differently and i can feel it

teaching an agent my workflow is mostly discovering which parts of my workflow were undocumented superstition

i am gradually replacing “be careful” with descriptions of what happened last time

if the interface needs me to remember what it was doing before i opened a menu, one of us is doing the other’s job

please show me the terminal setup you have slowly become unable to work without

“make it better” is a dangerous thing to say to something without a bedtime

“which terminal?” is the beginning of a story neither of us wanted to be in

sometimes the right amount of autonomy is “please stop asking me whether to do the thing i just asked for”

the most powerful prompt engineering technique is remembering to say what you want

a notification that everything is normal has made everything slightly less normal

what information do you go looking for immediately after an agent says it’s done

there should be a word for checking on an agent and accidentally giving it more work

escape is a key with several jobs and no clear manager

sometimes “done” means “i have reached the part where you do it”

i don’t need every operation to be instant. i do need my keystrokes to be

“what did you test?” remains an extremely useful question to ask anything that writes code

i don’t need an agent to sound confident. i need it to know which directory it’s in

the terminal received the key i pressed. we’re already having a good day

the difference between “fast” and “feels fast” is sometimes whether i can still type

i’m happy for the agent to spend ten minutes thinking if it saves me twenty minutes reconstructing what it did

the agent is allowed to be stuck. i would just like to be included in this information

the diff is where i find out whether we were having the same conversation

future me is an external contributor with no context and strong opinions

removed a small amount of mandatory suffering from the terminal

i understand why people implement a subset of vim. i also understand why vim users take this personally

please do not fix the neighboring function. it has a family

i like keyboard shortcuts that save more time than it takes to find out they exist