Notes on what works.
Short pieces on PLG, AI-native growth, and the unglamorous mechanics that move products forward.
-
AI-Led Growth: The Motion After Product-Led
Half the industry says PLG is dead because the next user of your software will be an agent. Both sides are arguing about the wrong layer. The economics of PLG survive. The assumptions of its funnel do not.
-
The Company Operating System
A persistent agent that holds what the company knows and gets asked before a person does. Context coverage as a tracked metric, why summarizing transcripts costs you recall, and the agent as gatekeeper.
-
The Cheapest Token Never Arrives
Spotify routed bulk file reads to a cheaper model and reported around 90% token savings. The number is workload-shaped. The finding underneath it, that soft rules do not hold on agents, is not.
-
Agent Teams for a Product Org
An agent team is a permission model with names attached. Which subagents a product org actually needs, what each one is allowed to touch, and the handoff where quality quietly dies.
-
Automate the Handoff, Not the Task
Most AI use-case lists are task lists, and tasks are not where the time goes. What to automate for support, sales, finance, R&D, product and leadership, and why access is the real constraint.
-
Skills Are Institutional Memory
A prompt is typed once. A skill is how your company does the work, written down so an agent can execute it. What belongs in one, what rots, and why the first draft is a process audit.
-
Context Engineering for PMs
What enters the context window decides what your product can be right about. The 4 strategies, the failure mode of each, and why retrieval is confident about absence.
-
Buy the Agent or Build the Grader
Build versus buy is the wrong frame. The question is who owns the pass bar, why the framework choice matters less than anyone thinks, and the wrap option nobody puts on a slide.
-
What an AI Automation Is Actually Worth
The same build, 3 prices, and the reason is not leverage. Time removed from a queue, what that time costs, how often it runs, and the review burden nobody quotes.
-
When Your Next User Is an Agent
Activation was built on a person in an interface. API-first products bent that model. Agent-first products break it: the user is software with a goal, a tool registry, and no patience at all. 3 users, 3 funnels, and no second chance.
-
Production Is an Org Problem
Almost every company has an agent pilot. Far fewer have one in production. The published surveys disagree on the spread and agree on the causes, and model quality is not among them.
-
The Agentic Growth Team
Growth teams spent a decade automating the funnel. What is being automated now is the loop itself, and the failure modes have recognizable shapes: experiment soup, metric gaming, and an accountability gap nobody staffed.
-
Evals Are the New QA
Every AI product works in the demo. Production is a stranger asking about their own data with money on the line. The scenario library is the new PRD, production data is the only honest test set, and the eval suite belongs in the release path, not in a report.
-
Who Grades the Grader
Once a suite outgrows hand review, a model grades the model. That judge fails the same way the product does, confidently and fluently, and almost nobody measures it against a human first.
-
You Can Only Loop What You Can Grade
Anthropic's Claude Code team sorts loops into 4 types. Read the table and it is not a capability ladder, it is a delegation ladder: the check, the stop condition, the trigger, the prompt.
-
When Resolved Becomes a Price
Bill per resolved outcome and the rubric deciding what counts as resolved stops being a quality artifact. It becomes the billing engine, and an eval failure becomes an invoice that never goes out.
-
Pricing After Seats: What Growth Teams Own Now
The human seat was the unit of SaaS pricing for 2 decades. AI agents broke it. When the unit of work detaches from the unit of pricing, the pricing page becomes a product surface, and that makes it growth's problem.
-
The Free Tier Meets the Inference Bill
Free was a marketing expense that rounded to zero. Then every feature acquired a model call and every free action acquired a real cost. Free survives, but now it needs a budget and a conversion job.
-
The AI Trust Gap: Why Traditional Onboarding Fails AI Products
The first session of an AI product is not a tour. It is a trial. The user is deciding whether the model is worth their attention, and most onboarding flows answer the wrong question.
-
Friction Was Doing a Job
In-assistant buying removed every step between intent and purchase, and converted worse than sending the shopper to the merchant. Removing steps helps when the steps are overhead and hurts when the steps are the argument.
-
The Transparency Tax
Disclosing chatbot identity before a sales conversation cut purchase rates by 79.7% in a field experiment. The timing fix that recovered it is exactly what Article 50 of the EU AI Act removed on 2 August 2026, which makes the disclosure a product surface nobody is measuring.
-
The Moat That Exported Itself
Every deck calls memory the last AI moat. Then the 3 companies with the most of it to protect all shipped memory portability inside a month. Whatever survives a plain-language export was never a switching cost, and the residue is the part worth defending.
-
Attribution After the Click
Growth measurement was built on the click. Discovery moved inside LLM answers, where the click never happens. Share of answer, branded-search shadows, and honesty about what is no longer measurable.
-
The Compound Effect of Small Funnel Fixes
Growth teams chase 10% wins and miss the ones that compound. A 5% lift in 3 places does more for ARR than a single redesign, and it ships in a week.
-
Distribution Is a Product Skill Now
Product people were raised on the belief that great products sell themselves. Building got cheap, discovery compressed into AI answers, and distribution became a design input instead of a department.
-
Metrics That Move Teams, Not Just Dashboards
Most product dashboards measure what is easy to count. The metrics that actually change behavior are the ones a PM, designer, and R&D lead can argue about in the same room.
-
Growth Runs on the Warehouse Now
The growth stack was a chain of tools, each with its own copy of the customer and its own version of the truth. The copies are dying, and what a growth team owns has changed from tools to definitions.
-
Building Retention Before You Need It
Churn is a receipt, not an event. Retention is an instrumentation problem long before it is an analysis problem, and the cancellation flow is the highest-signal surface most products waste.
-
The Buyer, the Admin, and the User
B2B onboarding keeps failing for a reason UX polish cannot fix: the flow is designed for 1 person and 3 different people show up, each with a different definition of value. Activation is a chain, not an event.
-
Going Enterprise Without Killing Self-Serve
Moving upmarket is usually framed as outgrowing self-serve. Backwards. The enterprise deal now starts inside the self-serve product, and in 2026 the evaluator reading your surfaces might not even be human.
-
Time to First Dollar
A marketplace has 2 activation problems running on different clocks. The supply side activates on first earning, the demand side on first satisfying transaction, and referral loops built before either one works amplify churn.