Where did that customer first hear about MailCheap?

Domain availability cuts the shortlist harder than taste does. Most of the good two-word names are gone before we start.

TypeScript, because one language across browser and server saves a lot of context switching.

The check mostly scores mouse movement and browser fingerprint before the click lands. An agent in a real browser profile looks ordinary on both.

N+1 queries tend to be the first one that bites, since the page passes every test on ten rows and crawls at ten thousand.

PUT is idempotent because repeating it leaves the same final state. POST needs an idempotency key header, otherwise a retry after a timeout charges the customer twice.

I would run it through a long multi-step agent task before ranking it. Short benchmarks hide how a model drifts after fifty tool calls.

A model can sweep a whole codebase for one bug class in minutes. Old bugs that nobody found worth hunting for are the ones surfacing first.

A skill is a folder with a markdown file and a few scripts, so supporting it costs another tool very little. That low cost is probably why it spread so quickly.

The 42 year old usually already has a network of buyers to call on day one. The 22 year old has cheaper runway and can afford more wrong turns.

An API that already exists usually means the demand is proven. The gap is normally in the workflow built around it for one specific kind of user.

Resolution rate per conversation is a better target than first response time. A 3 second reply that needs a follow-up counts as two contacts for the customer.

Tests that fail loudly are what let an agent run unattended. Without a check after each step it drifts, and we only find out at the end.

Claude for long refactors that touch many files. It keeps following the repo's conventions late in the session, where other models start drifting.

Most of that burn is the model re-reading the whole codebase every session. A short written spec per module keeps each session's context small and the limit lasts much longer.

Codex for the coding work, since an agent in a terminal carries over directly. Porting the rules file and the habits would take the longest.

Picking the right concept gets easier after asking what happens when the component goes down. That question alone separates a queue from a cache for most designs.

What prompt did you use for the design test?

A useful success definition is one number measured before and after, such as minutes per support ticket. It also shows when a single agent has hit its limit and a second one is justified.

Monolith, and split out a service only when one part needs to scale or deploy on its own schedule.