The generative trio
On the possibilities of consensus between product, design and engineering tooling
The scintillating development of large language models is perturbing many landscapes. One of these landscapes is product management. In The fifth product risk I referred to product trios. An engineer-designer-PM unit discovering opportunities and delivering outcomes. A multi-disciplinary, autonomous team manipulating product lines.
It's not unreasonable to say that a primary bottleneck for their progress is communication. True, the engineer-designer-product trio is united. They are embedded in the same organisational culture. They have a shared mission. They use common socio-technical systems—or at least ones that interface. They develop consensus on the opportunity space in front of them. They collaborate to manifest outcomes. Yet, each has their own particular flavour of reality that they must transcend in order to perform.
In fact, each interaction between the trio—product <> engineer, engineer <> design, design <> product—can be modelled as an information channel, a lá Shannon's information theory:
"The heart of his theory is a simple but very general model of communication: A transmitter encodes information into a signal, which is corrupted by noise and then decoded by the receiver. Despite its simplicity, Shannon’s model incorporates two key insights: isolating the information and noise sources from the communication system to be designed, and modeling both of these sources probabilistically. He imagined the information source generating one of many possible messages to communicate, each of which had a certain probability. The probabilistic noise added further randomness for the receiver to disentangle."
Imagine a product trio ensorcelled within such a model. The communication amongst them must retain the fidelity of the information intended for transport whilst remaining protected against the noise it's exposed to during encoding, transmission and decoding. This is where LLM-landscape-perturbance comes in.
To see how, consider an author penning a novel. They write a line in English. Simultaneously, this line is auto-magically translated into Spanish and Mandarin. The author continues to unfold the story. Eventually, they generate equivalent novels in Spanish and Mandarin, alongside the English one. One author with native-grade command of a single language writes a novel in three.
Yes, I understand that translation is not as simple word-for-word mappings. However, bouncing a view of a product between the respective lenses of product management, design and engineering could be that simple.
Rob Haisfield recently provided a clue about how a piece of this dance may work. In an X thread about generating designs from text he said:
"My mock-ups are more like visual specs for the UI designer to make what goes into prod. I want to use AI to rapidly iterate through high fidelity versions of my low fidelity prototypes, to round out my skillset as a PM, product designer, & consultant. ....
Let me handle the copy, elements displayed, flow, and general layout. Let me put rectangles and text boxes on a screen.
I just need AI to zhuzh it up, even out spacing, adjust the icons, colors, fonts, etc."
The thread also includes some cool demos of generative design workflows. And I'm aware of tools out there—like Merge and TLDraw—that bridge development and design. But the current crop don't get to consensus between product, design and engineering. "Consensus" is the key term here. It's derived from computer science and distributed systems design. ChatGPT summarises it thus:
"Consensus in distributed systems is a critical challenge where multiple components (such as servers or nodes) must agree on a specific value or state, despite potential failures and the lack of synchronous communication. This process is vital for maintaining consistency and reliability in systems where components are geographically dispersed or have unreliable connections.
Key issues include ensuring fault tolerance, even in the presence of Byzantine failures where nodes behave unpredictably; managing asynchronous communication where there is no guarantee of message delivery times; and achieving agreement, termination, and validity.
These challenges are addressed through various consensus algorithms like Paxos, Raft, and Byzantine Fault Tolerance (BFT) mechanisms, each balancing trade-offs between complexity, speed, and fault tolerance. The consensus problem is fundamental to the functioning of diverse applications, from distributed databases and network systems to blockchain technologies."
My thesis is simple. It will soon be possible—given the emerging tooling and the innovation that's driving it—to simultaneously bootstrap product, design and engineering systems with only basic competence in one of the three domains. Heck, potentially with expertise in none of them. Whether one or none, it will soon be possible to go from, for example:
A button that a website visitor clicks to
A simple, lovable, complete product to
A shiny, sharp, value-generating-at-scale solution
The way to walk that road is consensus.
Take fundamental, low-to-mid resolution structures from each of the domains. From engineering, take pseudo-code. From design, take wireframes. From product, take user stories and acceptance criteria. Now, establish consensus between each in a way that propagates changes in one to its parallels. Finally, add in per-domain generative tooling that allows one to increase scope, scale and complexity.
The result? Perhaps I move a boundary in a wireframe which alters an acceptance criteria which adjusts a React component. Perhaps I refactor an authentication integration which results in an extra step in the user's interaction journey which generates equivalent user stories and acceptance criteria which...
I suspect you get a whiff of why the establishment of this consensus matters. And if you do, then the call to build takes on a new aspect. The barriers to entry have—and will continue—to collapse. At least in the world of bits—hardware remains pretty hard. But it seems that product creation is on the cusp of a "creator moment". And it's thanks to the emerging consensus—enabled by LLMs—between the product trio.