← home

LET PEOPLE ASK THE QUESTION IN PLAIN WORDS

ok here's a thing that bugged me for ages. a super simple question, like "what's going
on with this customer?", had this slow, clunky, human-shaped answer. someone copies an
email, goes hunting for an engineer, the engineer runs a query, screenshots a row,
comes back... and by then the customer's already given up and gone.

so I spent a couple of evenings building a way to just... ask. now anyone on the team,
in plain english, can find out what happened to someone's account and get a real answer
back. no pulling an engineer out of deep focus. (engineers really, really do not like
being pulled out of deep focus, ask me how I know lol.)

two ideas carried the whole thing honestly.

first one: knowledge and power should be separate. the tool does the doing. a little
written guide carries the meaning. on its own the assistant would just kinda guess at
everything. with the guide it actually knows our words, our common situations, the
right way to look stuff up. and the guide is where the real value lives, so I tried to
write it like docs a new teammate could learn from, not like some clever magic prompt.

second one, and this is the spicy one: the safest way to hand power out is to wall it
in really tight and then just... trust the walls. I gave it the ability to look, and
only to look. nothing it does can change anything. a runaway question just stops itself
after a few seconds. once the walls are real, you can actually relax inside them.

and the part I'm weirdly proudest of isn't the tool at all. it's one tiny
non-negotiable rule: whoever changes how the system works has to update the guide in the
same breath. because tools rot the instant the docs around them stop telling the truth.
that little bit of discipline is the whole reason the thing still works six months
later. 🌱

cost? a few bucks a month. saved? a small recurring tax on everyone's attention. yeah.
that trade is basically always worth it.


---

this blog runs on [pre](github.com/connect-kai/pre)