RZ · Portfolio · 2026

← Back home

Design note · The guide cursor

Point, don’t tell

When you're lost in a product, help shows up three ways today. A wall of FAQ text. A support bot that types “go to Settings → Privacy → turn off Usage Data” and leaves you to hunt for it. Or an AI agent that takes the wheel and does it for you — so you never learn where anything is, and you're not the one who pressed the button.

I wanted a fourth. An assistant that doesn't talk at you from a side panel, but inhabits the one thing everyone already knows how to read — the cursor. You ask; it travels to the actual control and either explains it in place or rings it and says “this is it — you press it.” Because its answer is always a real spot on the screen, it can't hallucinate a location the way a chatbot can. And because it points instead of acts, you stay in control.

Meridian CloudOverview

Overview

us-east-1
Requests / 24 h2,847,012 4.2%
Error rate0.12% 0.03%
P95 latency94 ms 2 ms
Active tokens3--
Throughput (24h)
Resources
Environments3us-east-1
API Keys3us-east-1
Team members4--
Log volume1.2 TBus-east-1
Recent activity
API key rotatedryan.runsheng03:14
New member invitedsys-deploy02:07
Environment deployedci-pipeline01:52
Usage limit 80%system00:31
Billing updatedryan.runsheng-1d 18:04
SSO domain verifiedryan.runsheng-1d 09:11

What I decided, and why

A cursor, not a chatbox
Everyone already knows how to read a cursor. Borrow it and the help explains itself — you watch it go to the thing and stop. No paragraph to read and translate back into clicks. It doesn't teleport, either: you see it decide and travel, so where the answer lives is something you watch, not something you parse.
Point, don't do
It guides; you act. The opposite of an agent that takes the wheel — you keep your hands on your own product, you learn where things live, and for anything irreversible you're the one who commits.
Grounded, so it can't lie about where
It can only point at elements that are really on the screen. The answer is a place, not a claim — so it physically can't send you to a button that doesn't exist, the way a chatbot confidently can.
Answer or guide — never a wall of text
Two moves only: pop a one-line answer beside the thing (“what is this?”), or travel to it and ring it (“where do I…?”). The agent picks the move; you never parse a reply.