The Spec Is Not the Handover: Intent, AI-DLC and the Return of the Room

In The Compass You Never Handed Out, I made the case that scope creep is rarely a scope problem. It is a compass problem. Teams ask for more detail when what they are missing is intent.
The same mistake is now happening at AI speed. This time it is called spec-driven development.
Spec-driven does not mean waterfall
Many organisations hear spec-driven development and picture a familiar sequence. Product writes a complete, detailed specification. It gets handed over. Development builds it.
That is waterfall with a new label, and it misses the point of AI-DLC entirely. The intent behind AI-DLC is to put people together, shift product and domain expertise left, and have them work side by side every day.
The reason is human. We make sense of the world in a linear way, step by step, assuming a predefined plan leads to a predictable outcome. Much of business and leadership still runs on that mindset. Spec-driven sounds like it confirms it.
In reality, AI-DLC does the opposite. When people, agents and decisions connect in tighter circuits, speed increases until the work is no longer linear. At a quantum level, there is no distinction between past and future. Something similar happens in a bolt. Specifying, deciding and building no longer sit in separate phases. They happen in the same moment. The linear mindset is not wrong, it is simply too slow for what is now possible.
Look at what spec-driven actually means. Instead of working in code files, people work in the spec. It becomes the source of truth, and the agents turn it into code. If the spec is where the work happens, it cannot be a document that is finished and passed on. It has to stay alive for as long as the work does.
The spec is not a gateway between product and development. It is a shared artefact that evolves around the intent brought into the room.
The principles still hold
The Agile Manifesto turns 25 this year. Individuals and interactions over processes and tools. Customer collaboration over contract negotiation. Responding to change over following a plan.
AI-DLC does not replace any of this. It changes the cost of building. When building becomes fast and cheap, the bottleneck moves to how clear and how shared the intent is.
AI does not change the principles. It removes the excuses. Get intent wrong, and AI will help you build the wrong thing faster than ever.
What teams really need from product
Teams often say they depend on product for detailed specifications. They do not. They depend on intent. When the why is clear, the detail can be shaped by the people doing the work.
Product’s job does not change. Discovery stays the same. Product still explores, talks to users, understands the problem and matures a PRD, often in collaboration.
What changes is the purpose of the PRD. It is not there to answer every question in advance. It exists to carry three things into the room:
- The intent. Why are we doing this?
- The preferred outcome. What does success look like?
- The workflows users will be working within. How do they move through the problem?
A PRD should carry mental models, not instructions, and well-formed outcomes let a team model success before building begins. In AI-DLC, both become essential.
Intent is not owned by product alone. Engineering and design must bring intent around the technical architecture and the design. Everyone who holds a piece of the picture is responsible for bringing it.
Who asks and who decides
A bolt in AI-DLC is a short, focused cycle, typically one to three people working with AI agents.
In Scrum, the team breaks work into stories as tickets, separate from the code. In a bolt, stories become markdown files in the repository, next to the code, ready for the agents to build from.
The common mistake is to believe humans must now write every detail so the AI can execute. That is backwards. The AI asks the questions, based on the intent. The people in the bolt answer them and make the decisions. The AI surfaces the ambiguity. Humans own the judgement.
The spec is not an input to the collaboration. It is a record of it, growing with every question answered and every decision made.
Google’s recently open-sourced Elephant-Goldfish Model for spec-driven development makes the same point from another angle.
The elephant is a deep-context agent. It remembers the history, the architecture and the user journeys, and it drafts the spec.
The goldfish remembers nothing. It reads only the spec and the code, and tests whether the design holds up.
A goldfish is the perfect test of whether intent travelled. If a reader with no context cannot find the why, the intent stayed trapped with whoever wrote it.
This is not only true for agents. In every organisation, many people work like goldfish, acting on what is written and said. The question is whether we hand them the compass.
Where product gets back into the room
A popular belief says AI makes Agile ceremonies obsolete. No more daily stand-ups. The opposite is true.
Bolts start features that depend on other features, and those dependencies must be resolved. Product cannot sit in every bolt. In practice, product prepares with their own agents, often on the first day, and then the people are no longer in the same room.
The idea that product, with every decision, is always in the room is wishful thinking.
Science has known this for decades. In the 1970s, Thomas Allen at MIT showed that communication between engineers drops sharply with physical distance. Beyond roughly 50 metres, regular communication almost stops. This became known as the Allen curve. Distance does not have to be large to break the flow of intent. A different floor, a different calendar, a different bolt, or simply a different focus of attention is often enough.
Neuroscience adds another layer. Shared attention, people attending to the same thing at the same time, is increasingly recognised as a foundation for focus and collective performance. Which means the distance is not only physical. Two people can sit a metre apart, and if their attention is not shared, the distance reappears. The Allen curve applies to the mind as much as to the floor plan. Mental distance breaks the flow of intent just as surely as physical distance does. The stand-up is one of the few moments where attention is shared by design.
The daily stand-up is how product gets back into the room. It keeps decisions down to hours, not the weeks we accepted before Agile and again under scaled Agile. The AI raises the questions, the stand-up brings them to whoever holds the intent, and the decision flows back into the spec the same day.
This is not theory. It is a pattern that shows up quickly once bolts start running. The AI raises questions about edge cases and dependencies, and no one can answer them. Not because the detail is missing, but because the intent was never worked through before the bolt began.
When that happens, the stand-up is what prevents a full stall. It is where the gap becomes visible, and where the decision is made to pivot and replan the bolt. Without that daily touchpoint, teams keep building in the wrong direction, or keep waiting. Either way, time is lost.
Fit the process to the reality
There is no single right way to run AI-DLC, and there should not be.
Being truly agile means adjusting the process to the reality of the organisation. For some, bolts within sprints are the right fit. For others, a Kanban way of working serves better, where work flows continuously and blockers surface the moment they appear.
The deciding factor is what supports the business and its daily processes. Copying a framework by the book is not agility.
For leaders and executives, this means adjusting expectations. AI-DLC is not a way to remove people from the room. It is the reason the room matters more. The human factor, the communication and the mindset are what make it work.
Bring the compass into the room
The mindset that must go is the belief that everything can be specified perfectly before work begins.
Product must still have explored. Product, engineering and design must still bring intent around the problem, the architecture and the design. But the spec is not where the conversation ends. It is where it begins.
The next time a bolt starts, look at what is brought into the room. A stack of detail, or a compass? Then look at what the AI asks, and who is there to answer.
The spec was never meant to be the handover. It is the conversation, written down.

