Writing
On AI-first product development — from a practitioner who built a framework to solve it.
What product development actually requires
Structural problems with how product organizations work today — and the disciplines that address them.
The Translation Tax
Why AI coding agents make product management harder before they make it easier — and what changes when you remove the translation overhead from the PM's plate.
The Demo Gap
Why what stakeholders see and what engineering builds are never the same thing — and why that's a structural problem, not a communication one.
The Two Cards
A hypothesis and an experiment are not the same document. Keeping them separate is not a formality — it's the thing that makes hypothesis-driven development actually work instead of generating paperwork.
Why Your OKRs Don't Work: The Baseline Problem
Most OKR failures aren't format problems — they're baseline problems. The team set a target without knowing the current number. Here's why that gap is structural, and where in the process to close it.
The PM Who Never Wrote a Test
AI can generate acceptance criteria in Given/When/Then format from a PM's intent. The generation is easy. What it actually takes for an org to consume it — and why that's a different kind of work entirely.
Acceptance Criteria Are the API
The handoff from product to engineering is the acceptance criteria — but a bare list is only the signature of the API, not the contract. Why the discovery context behind the criteria is what gets the right thing built, and why the repo beats the ticket for carrying it across the seam.
Where the role is going
Forward-looking questions about product management as the tools and the stakes change. Not settled answers.
The Debt AI Can't Refactor
AI made refactoring cheap, so senior engineers are calling technical debt defused. They're half right. Code debt got cheap; decision debt — the unrecorded "why" behind a system — got worse, because cheap code stopped forcing the decisions that used to come for free. The debt refactoring can't reach, and what to do about it.
Why Engineers Should Want PMs in the Repo
The viral demos say non-technical PMs are building products "without engineers." The framing is backwards. PMs in the repo isn't a threat to engineering — it's a promotion, because the slop risk makes the verification gate and the machine that builds the machine the highest-leverage work, and that work is engineering.
The Framework That Forgot Product
A widely-shared AWS talk lays out a four-question operating model for the agentic era — economics, talent, structure, governance — and never once says the word "product," even though AWS's own written guidance names product as a protected role. How the function vanishes the moment the org chart gets drawn, and why it's load-bearing in all four pillars.
When the Product Is the Agent
The AI-First SDLC solves for building software with AI. What happens when the software you're building is AI? The questions PMs will need to answer don't have established frameworks yet.
When Execution Becomes Abundant
Engineering leadership is pushing toward agent fleets that do 98% of development. If they deliver on that promise, the bottleneck doesn't disappear — it moves. Here's where it lands and what Product Orgs need to build before the fleet arrives.
The Absorption Wall
When Execution Becomes Abundant argued the bottleneck migrates from engineering to Product. It stopped one beat short. Even a Product org that fully closes that gap runs into a second constraint on the other side — one that isn't staffable, because it lives in the customer, not the org.
The Best Context Isn't the Most Context
The path to better agents isn't more context — it's the right context. Why enterprises need to start treating context as a product, right-sized and governed per agent, and why deleting old context matters as much as adding new.
The Best Context, In Practice: A Bot Named Chiefys
An update to The Best Context Isn't the Most Context. A real tool at customer.io — an internal bot called Chiefys — operationalizes the argument that context is a product and deprecation must be first-class. What it confirms, and what it sharpens.
You Can't Generate Alignment
Executives are the best slop-detectors alive, so AI-generated alignment structurally fails — and the next model won't fix it. The PM's surviving craft is decomposing the problem and calibrating it to the room, not generating the artifact.
The Unicorn the Org Chart Wants
Managing a high-agency PM, generating executive alignment, and operating a fleet of agents are three different jobs — and the agents take none of the first two off the leader's plate. Why product leadership splits into a human track and a platform track, and the operate-the-fleet role no org chart has staffed.
Agents Run the Line, Humans Hold the Gates
Ask how much of product discovery an agent can take over and you're measuring the wrong unit. Every node is labor plus a gate decision — agents are taking the labor across the line, but the three gates that decide everything stay human, because discovery has no test the way delivery does.
The Stopping Rule
Paradox of choice and satisficing were never about how many options you have — they were about what generating and comparing them cost. AI deleted that cost, and two mirror-image failures showed up in its place. What replaces the stopping rule scarcity used to write for you, and why the Discovery loop's gates already are one.
Fast, and Still Yours
A business hands its whole product-and-engineering apparatus to an AI platform — not just the build, but the decisions about what gets built next. Two offers ride in one pitch, and the good version keeps a named person holding every gate that still matters.
Not So Fast, Agents
Companion to Fast, and Still Yours. The taste-test, the reversibility scoring, the gate — all of it quietly assumed the product held still once approved. An agentic product doesn't. What has to change when what's being shipped is a judgment function, not deterministic code.
Three Pillars, No Fourth
Citi posted a job for Head of Applied AI & Agent Factory — a live specimen of the industry's "AI Factory" movement and the Forward Deployed Engineering boom, absorbed into one org chart. What the structure reveals, and the one function none of the postings name anyone to own.
The Backplane Isn't One Memory. It's Four.
The product backplane was defined as one version-controlled layer of shared context. Cognitive science and a three-decade-old organizational-memory framework both say that bundles together things that don't belong under one governance rule — and describe what four separate memory types actually look like.