Pachinko pachinko
Pachinko pachinko

design notes

Useful AI is an interface problem.

I was the PM for Clippy in Office 97, so I learned the lesson the hard way: useful AI is not just about avoiding cartoons (shocking, but true). It's about designing around probabilistic behavior: systems that are wrong, right, and surprising in uneven ways. And systems that promise intelligence. These are the design challenges and lessons behind Pachinko.

01

Every answer sounds generic.

AI is trained on the average, so its default answers are for a typical person. Not you. The fix hasn't changed in thirty years: understand the user's context and preferences.

open
02

It sounds confident when it's wrong.

AI does hard things well and easy things badly, with equal confidence. You can't tell from the prose which one you got. Test it on work you already know before trusting it with work you don't.

open
03

I don't know what to ask.

An empty prompt box hands you the interface's job: know what to ask, what context to include, how to phrase it. We replaced the command line for a reason. Selection and visible choices beat an empty rectangle.

open
04

The first answer never seems right.

A single answer is a single bet. We learned this on the Answer Wizard: show a few good candidates and nobody cares which ranked first, only that the right one is there. A short list defers to your judgment, and raises both the odds and the feeling of being right.

open
05

Its help gets in my way.

The Clippy lesson: even good help costs attention when it interrupts, and bad help makes you work to get rid of it.

open
06

Can I keep some things private?

Yes - when privacy is a control you can see. The AI reads what you chose, the interface shows what's in scope, and everything you fence off stays fenced. Anything less, and users will never hand it real work.

open
07

The longer we talk, the dumber it gets.

Old goals, dead plans, and stale facts stay in the conversation and keep steering answers after they stop being true, and it's hard to see which ones. More memory isn't the fix. Sharper selection and cheap, fresh starts are.

open
08

I rebuild the same prompt every time.

A good result takes real setup: the goal, the constraints, the tone, the examples. Then the chat ends and the tool throws it all away. Save the setup once and run new material through it.

open
09

Checking its work takes longer than doing it.

A generated image is easy to reject; a thousand pasted numbers with five quiet errors are not. Automation only pays when mistakes show themselves cheaply.

open
10

I can't trust it to run alone.

An agent that runs unwatched compounds its mistakes, and you can't hand responsibility to something that can't take it. Keep the goals, the checkpoints, and the final say. The machine brings scale. You bring the judgment.

open

Pachinko makes AI personal and powerful.

Pachinko turns your context and rules into reusable instructions that can run repeatedly whenever new information arrives.

Pachinko is independent and not affiliated with Microsoft. Clippy, Office, Answer Wizard, Conversagent, and Microsoft product names are Microsoft trademarks.