Thinking
Notes on product decisions, AI, software systems and teams
Notes from decisions, problems and experiments I keep coming back to in Product and software work.
Each note stands on its own, but together they trace recurring questions about how teams choose, build and learn over time.
Editorial process
How I work with AI.
I have never been short of ideas, opinions, or things I wanted to try. What I often lacked was the time to develop, test, and connect them.
AI is changing that. I use it as a research, thinking, and building partner. Sometimes it helps me challenge an argument; sometimes it helps me turn an idea into something I can actually test and make part of how I work. What appears here always reflects my experience, judgment, and final editorial decisions.
Browse all notesStart here
Selected notes
Three notes on product decisions, AI-assisted development, and how teams interpret the same context.
- Waiting is one of the most underrated product decisions Product teams are often rewarded for shipping new things. But building something is not always progress. Sometimes the smartest decision is simply to wait.
- Designing for unattended development Unattended development is not the same as leaving an agent alone. It requires explicit eligibility, isolation, verification, stop conditions, and a separate decision about what may enter the stable codebase.
- Shared context is not shared understanding AI makes context easier to retrieve, but access to the same information does not automatically create shared understanding, alignment, or better decisions.
- Featured series Product Judgment in Practice Evidence, trade-offs, urgency, timing and disagreement before implementation begins.
- Building My Own AI Operating System A first-person series about moving from one general assistant to smaller AI-first services with explicit jobs, boundaries and ownership.
Questions I’m exploring
Follow a question through the notes.
These are curated paths through related observations, experience and external ideas. They are a better place to begin when a topic label is too broad.
- How do teams make better decisions? How information, trade-offs and timing shape product decisions when several directions look reasonable.
- How do people build shared understanding? How teams interpret context, surface disagreement and work across Product & Engineering boundaries.
- What changes when AI becomes part of the work? What AI changes about contribution, ownership and system design when it becomes part of everyday work.
Recent thinking
The latest notes.
The newest additions across product decisions, AI and work, software systems and collaboration.
- Adding MCP doesn't make a product agent-first Adding MCP to an existing API can create enormous value by making an established product accessible to agents. Designing a product with agents as first-class users is a different problem. AI and automation Software systems
- Better output makes shallow review more dangerous AI can make work look ready before its reasoning deserves trust. Review depth should follow the consequences of the decision, not the polish of the output. AI and automation Teams and collaboration
- Stop asking people for information the system already has Much process work moves facts between tools. Useful automation connects what is already recorded while keeping system interpretations separate from source truth and human judgment. Software systems
All notes
Showing all 20 notes
- Adding MCP doesn't make a product agent-first Adding MCP to an existing API can create enormous value by making an established product accessible to agents. Designing a product with agents as first-class users is a different problem. AI and automation Software systems
- Better output makes shallow review more dangerous AI can make work look ready before its reasoning deserves trust. Review depth should follow the consequences of the decision, not the polish of the output. AI and automation Teams and collaboration
- Stop asking people for information the system already has Much process work moves facts between tools. Useful automation connects what is already recorded while keeping system interpretations separate from source truth and human judgment. Software systems
- I need my AI dashboard to leave things out A useful AI dashboard should reduce the first pass over the work, not reproduce every open item. Its omissions, priorities, and explanations all need to remain inspectable. AI and automation Software systems
- Friday connects the services without owning their work A coordinating layer can compare state, explain disagreement, and direct attention without becoming the source of truth for the services it connects. AI and automation Software systems
- Designing for unattended development Unattended development is not the same as leaving an agent alone. It requires explicit eligibility, isolation, verification, stop conditions, and a separate decision about what may enter the stable codebase. AI and automation Software systems
- Why I started building Friday Friday is my second attempt at the original Jarvis ambition. It began when August and March created a concrete coordination problem between them. AI and automation Software systems
- I built March to plan with AI without becoming a content machine I wanted a view of the publishing runway without turning empty space into pressure. March keeps the plan visible so AI can help question it while the editorial decisions remain mine. AI and automation Product decisions
- I built August because copy and paste was not collaboration AI could help with a document, but the work around each answer kept falling apart. I built August to keep the draft, its review state and my decisions in one place. AI and automation Teams and collaboration
- I stopped trying to build Jarvis I started with a general AI assistant and no recurring job for it. The useful parts appeared only after I built smaller services around work I was already doing. AI and automation Software systems
- The transition to Product Management starts before the title changes The transition into Product Management often begins before the role changes, when attention shifts from implementation alone to problems, trade-offs, communication, and product judgment. Product decisions Teams and collaboration
- Shared context is not shared understanding AI makes context easier to retrieve, but access to the same information does not automatically create shared understanding, alignment, or better decisions. Teams and collaboration AI and automation
- AI accelerates contribution, not mastery AI makes it possible to contribute much earlier in a new environment. But contributing faster is not the same as deeply understanding the system. AI and automation Teams and collaboration
- Temporary solutions become permanent Temporary solutions are often necessary, but once adopted they become part of the product and can quietly turn into long-term constraints. Software systems Product decisions
- The urgency of customer requests Not everything that feels urgent is a real priority. In product work, reacting before urgency is understood can lead teams to the wrong decisions. Product decisions
- Product decisions are mostly trade-offs Product discussions often look like debates about the right solution. In reality they are usually debates about which trade-off a team is willing to accept. Product decisions
- Most product disagreements come from missing information Product discussions often look like disagreements of opinion. In reality, teams are often missing the same piece of information. Product decisions Teams and collaboration
- Waiting is one of the most underrated product decisions Product teams are often rewarded for shipping new things. But building something is not always progress. Sometimes the smartest decision is simply to wait. Product decisions
- The real job of a Product Manager is managing disagreements Much of product work is about helping smart people who see the problem differently find a way forward. Teams and collaboration Product decisions
- When should Machine Learning be used? Machine Learning should only be used when there is enough meaningful data available. Without the right data, a rule engine is often the smarter choice. AI and automation Software systems