Product & Design

Using Dot made me realize Roleward doesn’t need to become yet another Agent—what it should be is Agent-ready, so outside Agents can call its core capabilities safely and clearly.

The difference between Projects and Spaces isn’t whether you can save materials or collaborate with others—it’s the entry point for work: a Project centers on Chat, a Space centers on outputs like Pages.

This day was a video on RedNote.

Having put gains and costs together, I didn’t find an “average enterprise AI ROI” worth believing: the economics depend heavily on the task, baseline, users and workflow, and the better question is under what conditions a specific workflow is worth scaling up.

API prices are easy to look up, but they aren’t the whole bill: getting AI running, making its results trustworthy, and keeping it alive long term each carry costs you can’t see in the token price, and which is biggest depends on the specific workflow.

AI’s productivity gains on specific tasks like customer service and writing code already have decent real-world evidence, but between “one person working faster” and “the company actually getting economic value” there are still several steps.

A metric is meant to help me see the goal, but once I treat it as the goal — “this post has to get so many likes,” “apply to 5 jobs a week” — my behavior follows the number, and may end up further from what I actually want.

My product was ugly because design came in too late. While Codex was writing code it was already deciding the UI/UX for me; each page made sense on its own, but together they didn’t look like one product.

Once AI writes the PRD, generation is no longer the bottleneck. The new bottleneck is whether I can actually spot the changes, assumptions and problems in a document that’s fast, long and reads smoothly.

When building products with AI, what really needs distinguishing isn’t who wrote which document, but whether a piece of work is organizing and exploring, or deciding on the product’s behalf. AI widens the options and fills in detail; the PM converges, makes trade-offs and confirms.

The DESIGN.md Claude Design wrote for me had colors, fonts, spacing and components, and was still off: it explained what the interface should look like, not why the product should be designed that way.

Implementing harness engineering really is engineering-heavy, but once an agent can read data, call tools and act for users, the decisions about information, permissions, boundaries, acceptance and human intervention were always part of product design.

The hard part of AI products isn’t only bringing AI in; it’s managing its uncertainty. Eval design answers exactly “what counts as right, which errors must be stopped, and how the next version proves it’s better,” so it’s a core skill for how an AI PM manages quality.

I took Asterline, a demo I built, through the whole Eval Design process. The company and the data are fake, but the iteration loop — from Golden Set and Rubric to Failure Types — is real.

I used to think eval design meant running a few cases and giving AI a score. Now I think what it really has to answer is: for a specific task, what counts as right, what counts as wrong, and what to fix when it’s wrong.

Vibe coding has made “building it” cheap, but the difficulty of a product hasn’t gone away; it has moved later. A demo proves something is possible; a product has to prove it keeps creating value when inputs get messy and users don’t behave as expected.

AI makes “building it” easier and easier, and also lets us build things nobody needs faster. The Mom Test reminds me not to ask users “would you use this,” but when they last actually ran into the problem.

Rereading fast.ai’s Build to Last eight months on, my conclusion is: the better AI gets at writing code, the less we can give up understanding our systems. Code keeps getting cheaper; judgment keeps getting more expensive.

After reading AI in Design 2026, I think the question isn’t whether AI will replace designers, but that design is shifting from delivering design files to organizing a process that runs from idea to implementation.

I used GPT image 2.0 to imitate the illustration style of Airbnb’s city invitation cards. The images really are impressive, but the surprise feels predictable, as if something like a “soul” were missing.

My way of getting good-looking design out of AI is to break the job into steps: general agents set the framework early, design agents do the UI/UX in the middle, coding agents build it at the end, and each step gets the tool best suited to it.

Figma released its own design agent, and I’m curious whether its target users can move beyond designers to the wider builder group.

Of the many design tools I’ve tried, Claude Design turned in one of the more satisfying results. It surprised me more than Google Stitch did, though maybe I just hadn’t found the right way to use Stitch.

After going through 90 products from the Product Hunt leaderboards for March to May, I see the AI buzz shifting from “models and platforms” to “workflows and execution”: people don’t want a smarter AI, they want a product that finishes a piece of the task for them.