I built Considered because I kept running into the same problem with AI-generated interfaces.

I’d ask an agent for a screen and get all the right ingredients: cards, charts, a sidebar, decent styling. Everything rendered. But nothing felt decided.

The biggest number was often just whatever came first in the data. Actions had no real hierarchy. The interface looked designed, but there wasn’t much evidence that the agent had thought through what the user was actually there to do.

Considered adds a step before the build.

The agent has to define what decision the screen is for, what deserves attention, and which actions belong at the page, section, or object level. Then a second agent reviews the result without seeing the first agent’s reasoning.

A build can’t sign off on itself.

I ran two blind evaluations against a baseline, and Considered won 24 of 30 pairwise comparisons. The reviewers were models, not people, so I see that as a signal, not proof.

The more interesting finding was where it failed.

In the first eval, Considered invented more details than the baseline: names, policies, durations, things nobody had provided. More structure gave the agent more places to confidently fill in the blanks.

I added a ruleset to address that, and I’ve documented the failure alongside the wins.

Mostly, I’d love help finding the next thing that breaks.

If the direction fights your product, the hierarchy feels wrong, or the review approves something it obviously shouldn’t have, tell me. That usually means there’s a rule missing.