All Posts

A cheap prototype doesn't replace the conversation with users, it makes it better. But research can't be a phase at the start anymore.

A user story that is only a card becomes the prompt. Context, conversation, and confirmation have to travel with it, or the agent builds the hesitation.

A product can be tasteful and still make you hesitate before you hand it something that matters. Operators don't trust a lock icon. They trust honesty.

Models are extraordinary at preference. Judgment explains why something is the right choice, and thought is the care you can only put into a moment you have seen.

A product can work and still lose people after the trial. Fine is the new floor, and the hesitation is the part the model can't remove.

You don't need to become a developer. You need the working model you build by shipping small, real things before the skill names change again.

A sandbox repo is not only infrastructure. It is how agent feedback stays grounded in real components, real tokens, and real constraints.

Newsletters give you a map. A broken agent session gives you a working model of where judgment has to move.

A knowledge pipeline, an executive assistant, a consultancy offer, and a booking system for a training room in Colombia. Small builds, specific enough to break.

NIS2 is in force, and the people implementing it aren't compliance lawyers. How NIS2.chat and the NIS2INTEL tracker went from an idea to working products.

Agents fail users at the decision, not the execution. The decision gate is where the user's intent enters the loop, and where accountability stays.

The same five questions before any significant build. For products where an agent acts for a user, that conversation has to become a specification.

Clear requirements can still describe the system and miss the user. Intent engineering is design's oldest job, and you only learn it by building.

Prompting is tactics. Intent is strategy. The conversation before plan mode is the highest-leverage work in AI-assisted building.

The barrier to building dropped. The barrier to understanding what people need did not. Restraint is the taste decision AI cannot supply.

Linear won on a point of view, not on speed. Taste is judgment built from proximity, and both common responses to AI miss it.

Competent design is nearly free. The T-shaped designer got broader because owning the outcome now means owning the loop.

The source of truth moved from Figma to the codebase. Speed does not forgive bad taste. It amplifies it.

Directing agents in a production-shaped sandbox is not designers replacing engineers. It is where intent meets an outcome someone can evaluate.

The missing piece is not another Figma file. It is a 1:1 clone of production where design can stress-test intent against real constraints.

When the middle compresses, what is left is the front edge of intent and the back edge of evaluation. Three ways of standing there are defensible.

The middle of software work is collapsing. A lot of what designers called craft was execution, and the grief is about that discovery.

The consequential design decision is rarely visual hierarchy. It is where a capability sits inside someone's workflow, before Figma opens.

Consensus does not scale, process will not save a wrong direction, and design leadership that only advocates has already signaled that design is optional.

Waiting for alignment feels responsible. At agent speed it is how you ship a confident, well-built product in the wrong direction.

AI agents execute in an afternoon what used to take a team weeks. The beliefs I had about design leadership did not survive that speed.

Front-end work is forking. One path owns the system agents build with. The other owns the craft AI still can't evaluate. Both need intent.

AI made screens cheap and consistency expensive. The new front-end job is the interaction vocabulary that humans and agents both have to obey.

For PMs and strategic designers, throwaway code replaced the wireframe. The prototype is the conversation, not the product.

The barrier to a functional interface is gone. What remains is directing agents that are fast, chaotic, and good enough to change who gets to build.

Start with one problem and make every feedback loop faster. The tool matters less than the loop, and the sunk cost of exploring an idea is what changed.

Software is getting cheaper to build. What differentiates the work now is context, feedback, and knowing when the agent is confidently wrong.

NIS2.chat, a multi-agent design critic, and the internal tools at Omny. The real build is the system underneath the products.

A billion tokens later, the skill isn't vibe coding. It's directing intent, managing context, and knowing when to step in.