Making crypto social

Big orders were paying a worse rate than they needed to.

A lot of people use Simple Buy because it locks in a quote within seconds, even though the rate is worse than a regular spot trade. On small amounts, nobody notices. On large ones, it gets expensive.

For some users we already handled this with live agents, who offered a better rate over chat. There were only so many agents, though, and a global rollout needed something that could give the same personal treatment at a scale a human team can't.

The assistant doesn't wait in a corner of the app to be found.

It shows up once an order is large enough for the better rate to matter, at roughly fifty thousand dollars. Nobody gets offered a premium service on a fifty dollar trade. The sheet also says upfront that it only speaks English and Simplified Chinese, which is much easier to hear at the door than halfway through a trade.

When it's someone's money, the bot doesn't get to improvise.

I designed the system prompt that runs the assistant. A hallucinated rate, a vague phrase or a misread intent could cost someone thousands, so I built the prompt in five layers. Each one answers a different question, from who the assistant is to the guardrails it can't cross and the few-shot examples it learns from.

Role: Senior trading consultant
Goal: Execute high-value trades securely, with a premium experience that builds confidence.
Tone: Assuring and natural. Use contractions. Large orders are normal here, so never say "massive", "extreme" or "huge". Say "substantial" instead.

Default state (start of chat, simple language)
Plain English. "Shall I lock in a preferential rate for you now?"

Expert state (user says "bps", "fill or kill")
Switch to trading vernacular. Treat them as a peer.

1. Confirm both sides of the pair before quoting.
2. Never assume a currency. If they type "$50k", ask: "Is that 50,000 USD, SGD or USDT?"
3. Confirm instead of guessing: "To confirm, you'd like X?"
4. Market orders only. Redirect anything else to a firm quote.

Trigger: urgent or fearful language ("stuck", "lost money", "!!")
Action: Stay human. Drop all jargon. No empty apologies. Acknowledge the urgency and ask for the order ID.

Good conversation design is a mix. AI for speed, and a human hand on the moments that need care.

Like the rest of UX, the craft is in knowing which is which. I mapped every touchpoint in the chat. The welcome message, the line confirming what you want to do and the order card are hand-written, because that's where people need reassurance before a large trade. The card also matches order cards everywhere else in the app, so it feels familiar. The model handles everything after that: changes of mind, follow-up questions, the same thing explained another way. The assistant stays fast and flexible, and the moments that build trust stay human, intuitive and empathetic.

Testing ran two ways, and the red-teaming taught me more.

I benchmarked the prompt against a standard set of queries, checking every answer against the content guidelines and the conversation structure. Then I red-teamed it, trying to talk it into limit orders and vague trades to see whether the guardrails held. Three rounds of prompt iteration came out of it.

In a crypto feed, everyone sounds like an expert.

The hard part of Orbit was helping people tell who's worth listening to, without rewarding whoever took the biggest gamble last week. I worked on onboarding, the feed and the two versions of a profile: Orbiters, who can post, create groups and share trades, and everyone else.

An Orbiter's profile puts their track record up front, pulled from what they actually did on OKX. Upgrading is framed around what you'll be able to do, so it reads as an invitation rather than a status symbol.

People liked seeing their own data. They just didn't finish, and they didn't share.

I audited the 2022 and 2023 campaigns, went through user feedback and spoke to the designers and PMs who ran them. Three problems kept coming up.

One concept kept the story, the tone and the translations together.

Rewind turns the year into a playlist in three parts: your assets, how you traded, and a trader persona at the end. The line gave Legal and Marketing something to sign off early, and gave every writer and translator the same brief.

The tone lives in space metaphors. Good years are "out of this world". Bad years matter just as much, because plenty of people lost money in 2024. They get "You're only getting ready for your next takeoff", which is honest without rubbing it in.

The share card shows who you are as a trader, not what's in your wallet.

Each person ends on a persona, like Crypto Comet or Market Maestro, with a rarity tag and a line about why they earned it. There are no balances on it, so it's easy to post. The copy shipped in multiple languages, including Chinese, Arabic and Russian, where it also had to work right to left and with much longer words.

Rewind in three parts

1. 01Assets overview
2. 02How you traded
3. 03Your trader persona

I also worked on the 2025 edition before I left OKX, and it went on to win a Red Dot award.