I Killed an AI Feature Before Writing a Line of Code

The idea started as a one-line memory hint: "AI video summarizer." No spec, no context, just a note to self from an earlier session. It's the kind of idea that sounds appealing precisely because it's vague — vague ideas have no failure modes yet.

Round one: own-content summary

The first shape was a digest that watched dev-hub's own activity — new posts, new projects — and generated a short AI summary periodically. Under questioning, the first crack appeared: who reads this, and where? There was no answer that wasn't "maybe, eventually, an email." A feature with no confirmed reader is a feature with no confirmed problem.

Round two: on-demand or stored

Reframed again, as something a visitor could trigger on-demand instead of something running on a schedule. This solved the "who reads it" problem but introduced a budget problem: on-demand generation against a real model, on a personal site with no revenue, is a cost with no ceiling.

Where it collapsed

Once the no-budget constraint was on the table, the honest comparison was: the output of this feature — a title, a short body, a timestamp — is structurally identical to a Post. The only thing "AI digest" was adding was the generation step, not a new shape of content. Building a second model, a second dashboard screen, and a second public route to hold the same shape of data the Post model already holds isn't a feature, it's duplicated infrastructure with a novelty label on it.

Deferred, not built. The interview didn't produce a design doc — it produced a reason not to write one. That's the actual output structured pre-implementation questioning is for: catching "this is already something I have" before it becomes two things to maintain.