Re: Asking for help: Update German translation of NR using LLM

On further thought: are there "known good" states of the docs? By which I
mean commit X in the English is known to be a validated match of commit Y
in German (say).

If so I would tell the agent: look at the difference between HEAD and X in
English and apply that diff to Y in German. Then take a second agent and
tell it to review the work of the first, then take a human and tell him/her
to review the combined result.

It seems that for most pages this could be quite smooth (and there is this
thing that agents are increasingly good at catching each other's mistakes).

More importantly it focuses the time spent by the human onto where the work
is needed

And per David's point below, they are also improving in terms of trying to
be idiomatic and capturing the general feel of the conversation. The ones I
use never send me back humour, but they're not fooled by my sarcasm or
snarky comment: they see right through them and get to the meaning right
away (I poke at their ability to understand things on purpose). This seems
to be equally true on my work accounts (which use fairly powerful modes) as
well as my personal accounts (that I keep small)

L


On Sun, 26 Jul 2026, 14:16 David Kastrup, <[email protected]> wrote:

>  It was very "me".
>