Operating principles
Eight rules I keep coming back to. Each one has a receipt in the case studies.
Validate demand before engineering commits.
The cheapest experiment is the one that never gets built. Every product in the studio had a demand signal — a waitlist, a paid pilot, a usage pattern in an adjacent product — before it had a repo.
In practice — IQnition started as Study Sense, an AI chat product. Demand testing showed parents weren’t buying a chatbot; they were buying oversight. We pivoted the product, not the pitch. →
Define the metric before the feature.
If I can’t name the activation event and the retention window, the feature isn’t specified yet. Metrics are a design input, not a reporting output.
In practice — At ZTE the recommendation engine shipped with conversion and revenue-per-impression as the only roadmap criteria. Every A/B test on ad format and placement answered to those two numbers. →
Decouple the variable from the fixed.
Most platform leverage comes from finding the small part that changes per customer, per brand, per listing — and moving it out of the thing that is expensive to rebuild.
In practice — Sprint’s single-SKU program moved the carrier-brand surface out of firmware and into an OMA-DM payload: one build, one certification, every brand. Studio 3Sixty5’s media engine does the same thing to video — the motion is a template, the listing is the payload. →
Own the cost / quality / latency triangle.
An AI product’s unit economics are a product decision. I pick the models, the routing, and the fallbacks, because the P&L and the user experience are the same artifact.
In practice — Studio 3Sixty5 runs a multi-model image/video/text pipeline with a sub-10-minute target. Which model handles which step, and when a cheaper one is good enough, is a weekly roadmap conversation. →
Responsible AI is a product control, not a legal afterthought.
In regulated consumer markets — housing, children — the guardrail is part of the value proposition. Customers pay for output they can publish without a second review.
In practice — Fair Housing and MLS compliance checks sit inside the listing-copy generation path at Studio 3Sixty5; IQnition runs an LLM gateway with eval models and child-account constraints before any output reaches a student. →
Kill rate is a feature.
A portfolio only compounds if underperformers get deprecated. I publish the stop criteria at launch so the decision is mechanical, not emotional.
In practice — Of the five products shipped under PromptXchange Labs, each has explicit traction thresholds; the ones below them get sunset or folded into a stronger product. →
Write the narrative, then the roadmap.
Senior stakeholders fund stories about customers and economics, not Gantt charts. The roadmap is the appendix to a strategy memo that stands on its own.
In practice — Every product here has a one-page thesis that predates its first sprint. The case studies on this site are those memos, with the outcomes filled in. →
Be the PM who can read the diff.
I don’t need to write every line, but I need to understand the architecture well enough to make the trade-offs myself and to know when engineering is sandbagging or when I’m asking for something unreasonable.
In practice — Multi-agent pipelines, LLM gateways, CI/CD, mobile OS provisioning — the diagrams in the case studies are the ones I drew for the team, not reconstructed afterward. →
The lifecycle
PromptXchange Labs runs Validate → Prototype → Ship → Measure → Scale as a repeatable loop. It is explicitly a portfolio-management function: capital (engineering time) is allocated to the products showing traction and pulled from the ones that aren’t. The studio is one person’s answer to the question how do you run five products without a hundred-person org? — treat them as positions, not pets.