deltastring weekly
Zendesk news and incisive views
This week
I've been thinking about opinionated interfaces this week. The best tools don't just let you use them; they nudge you toward using them in the way that produces the best result. They teach habits as you go, and they pave the route worth walking.
Twitter is the one I keep coming back to. It started life as a group SMS system: short text fired off to a list. Then the users got hold of it. They invented hashtags so things could be found. They built image hosts so things could be shown. They wrote link shorteners so things would fit. The platform's job, eventually, was to look at where everyone was already cutting across the grass and lay the paving stones.
Sometimes the desire path is the right path and we should pave it. Other times there's a reason the existing route still matters, and the kindest move there is to plant a hedge: make the wrong way gently harder, while keeping the right way obvious. The first job of any good tool is making sure everyone gets where they're going. The second, almost as important, is making the route worth walking.
This week, Mercedes committed to bringing physical buttons back across the range, after the wall-to-wall hyperscreen dashboard was universally mocked. The screen-as-luxury cue had quietly flipped: a huge telly meant money in 2005, a huge touchscreen in 2025 means ordering at McDonald's or at the Argos collection point. Buttons aren't nostalgia, they're the right interface for things you operate without looking. The status signal moved while the design teams kept building toward the old one, and Mercedes have decided to pave the path back.
From Deltastring
I find writing pretty hard. I'm really surprised at how pleased I am to see my thoughts on the page. If you've ever been the one person doing two jobs, and one of them is Zendesk, my inbox is open.
What caught our attention
2 sources · industry · 2026-05-04
Retraining frontline staff to work alongside AI is the bit most vendor decks skip. Walmart will be the case study the rest of the industry quietly copies, or a cautionary tale!
2 sources · industry · 2026-05-04
Zero to nine figures of ARR in months. This valuation reflects a real underlying bet: that the next generation of support platforms won't be a big email inbox with another decade of additional channels and extensions bolted on top. We agree and this is why we are building Deltastring AI. The shared-inbox metaphor has had a very good run but is reaching a conclusion.
2 sources · general · 2026-05-04
Most "AI deployments" still measure activity, not outcomes. Easy-to-measure metrics like CSAT and tickets-over-time are backstops, not insights, and the businesses getting real value from AI are the ones who decided what "good" looked like before they bought the product. Measuring the data improves the data so put the number on a screen the whole team can see and watch it move. We did a video with more detail.
What we shipped
Zendesk configuration management
| Edit SLA targets directly inline to adjust policy metric targets without leaving the page |
| Routine security audit fixes in this round include rate limiting, email bomb protection, and open-redirect handling |
Deltastring AI
AI-native support platform
Deltastring AI replaces raw Claude inside Beacon. It's an in-product assistant that answers questions about your Zendesk setup using only your own approved content. Every answer is grounded in something you've authored and points clearly back to its source.
| Every answer cites its source, with each response grounded in canonical content you've approved and a clear pointer back to where it came from |
| Your content stays your content, entering retrieval only after you've explicitly approved it and staying inside the boundary of your own tenant |
| Knows your setup the way Beacon does, grounded in the same canonical view of your Zendesk configuration that Beacon already maintains, so the answers reflect your actual environment |
Worth your time
Marc Andreessen's super-prompt for the perfect AI (yes, that Marc Andreessen)
Andreessen posted his custom system prompt on Twitter this week. It's a dense block of instructions: be a world-class expert in all domains, never hallucinate, never validate premises, use explicit confidence levels, be argumentative and provocative, lead with the strongest counterargument, and so on. Lance Eliot's Forbes piece is worth a read on what actually happens when you try to instruction-stuff a model into being smart.
The bigger thing it reveals is harder to shake. One of the people writing the cheques for half the AI industry seems to be operating on magical thinking about how these systems actually work. If the capital backing the next wave of products has that level of understanding, how do we end up with a sane market for new ideas?
Got a response to anything here? Something else to share? Reply to this email.
Forwarded this? Subscribe here.
Deltastring Ltd · deltastring.com
Unsubscribe · View in browser
What happened in Zendesk this week
Platform changes, security updates, and honest opinions from someone who builds on Zendesk every day. One email, Fridays.
“This is the only newsletter I want in my inbox, a beacon in a world of spam”