The previous process
Double diamond taught the sequence. The handoff conveyor belt was the cost.
For years the map was Discover → Define → Develop → Deliver. Diverge into the problem, converge on a wedge, diverge into options, converge on a ship. It was the right spine for consequential work: problem space before solution space, evidence before polish.
What strained was treating those phases like stations on a conveyor: decks and artifacts passed down the line while learning waited for the next gate. The names still matter. The operating model had to evolve.
What each phase produced
Swipe sideways for each phase
Phase by phase
These are modes of work, not sequential gates.
I move between discovery, definition, development, and delivery continuously, using working software, user evidence, and production behavior to decide what happens next. High-velocity teams cross these modes in the same week: a prototype exposes a discovery question; production traces become research; evals become requirements; a guardrail failure changes the interaction.
Discover
The problem you think you have usually isn’t the problem.
Discovery tests assumptions before they’re expensive. What comes out of this phase is a clearer picture of who has the problem, how severe it is, and what the team is secretly assuming. A research report is a byproduct at best. Interviews, journey mapping, competitive context, hypothesis statements. Where it helps, lightweight prototypes so stakeholders and users react to working behavior instead of slides.
Teams that skip discovery don’t skip the questions. They answer them later, with more invested, and less room to be wrong.
Interviews and insights · User personas · Competitive analysis · User journey maps · Hypothesis statements
Define
A problem well-framed is half the solution.
Define turns signal into direction: a problem statement specific enough to design against, and alignment on what success means. The temptation is to jump to UI. The discipline is staying in the problem until the viable wedge is obvious, including when the right answer is to narrow scope or stop.
High velocity isn’t more output. It’s bets small enough to inspect, cheap enough to reverse, measurable enough to learn from, and safe enough to ship frequently. Workshops, storyboards, and low-fidelity prototypes test direction before build investment. The durable output is a problem the whole team can argue with, and a short list of what we are not building yet.
Collaboration artifacts · Storyboards · Low-fidelity prototypes · Problem framing · Small reversible bets
Develop
Ideas are cheap. Tested ideas are not.
Develop explores and winnows options. By the end, what matters is a handful of viable ideas with evidence behind them. Volume is easy. Interactive prototypes and validation before friction hardens. High fidelity waits for a confirmed direction, however much the room wants to feel progress. Prototypes earn their way into continuous delivery; most never should.
For AI products, prototypes test more than interface usability. They expose model behavior, tool use, uncertainty, handoffs, failure and recovery states, and the points where human judgment must remain in control: acceptable and unacceptable behavior, where autonomy is appropriate, and how the system escalates.
This is where the design-to-code loop shows up: clickable flows, staging-shaped builds, representative scenarios run repeatedly. AI compresses synthesis and iteration; it still doesn’t replace watching an expert user hit a wall you didn’t anticipate, or deciding most prototypes should be thrown away.
High-fidelity designs · Interactive prototypes · Staging-shaped builds · User validation · Behavioral evals · Failure and escalation states
Deliver
Prototypes earn their way into continuous delivery.
Deliver turns validated behavior into a production path: small increments, cheap to reverse, instrumented enough to learn from. Design stays embedded through implementation, testing, release, and iteration. No baton pass of screens and specs. Delivery stays a learning system rather than a release ceremony. Product requirements become testable scenarios and behavioral evals.
Production traces, user corrections, quality and safety signals, latency, and cost feed the next decision. Training and iteration paths are designed in from the start instead of bolted on after disappointment. The compounding work happens here: a mental model that deepens, a system that improves under real use, a product meaningfully different six months later because decisions were validated with evidence.
Design systems · Continuous delivery path · Behavioral evals · Production traces · Quality, safety, latency, and cost metrics · Feedback loops
What’s evolving
AI accelerates production; evidence still determines direction.
The job isn’t research → define → test → ship, with AI making each step faster. It’s the fastest trustworthy loop between a real problem, a working system, observable behavior, and the next decision, using AI to compress execution while strengthening evaluation, human judgment, and operational safeguards.
Vague decks bouncing through tools create drag. Durable intent (constraints, states, what good looks like, where autonomy stops) lets a team move from question to interactive test without losing the plot. A prototype orbit throws most ideas away on purpose. A product orbit is continuous delivery with learning built in: harden what survived, ship in small increments, instrument it, and feed discovery again.
Same craft. Different medium: intent close enough to code that agents and engineers can execute it, and learning that happens in fast, reversible loops instead of at the next departmental gate.
How the phases land now
The same four names, in continuous motion.
Discover and Define still frame intent. Develop runs the prototype orbit, including how a probabilistic system should behave under load. Deliver is the product orbit: prototypes that earned their keep enter a continuous delivery path, ship in small increments, observe, and loop back into discovery. The evolution isn’t renaming the work. It’s refusing to freeze learning between stations.
Framing
Discover and Define still earn the build.
-
Discover
Who has the problem. How severe · what we assume.
Feeds Intent
-
Define
Problem worth solving. Wedge · what we are not building.
Writes Intent
Orbits
Develop and Deliver keep learning in motion.
-
Develop
Interactive tests. Most ideas discarded.
Evidence for Intent
-
Deliver
Ship · instrument. Loop back to Discover.
Compounding
Product design for consequential systems.
The commercial shape of this method: 6–12 week embedded principal engagements.