nickintheloop

Everyone's posting "Claude edited my video" like it took one prompt. I don't know about others but for me It didn't. It took me a week to get a setup that works. Here's the whole thing as a skill, plus every tool I used: github.com/nstavr04/make-con… Feel free to help me improve it

I like the idea of getting useful work done before full autonomy. Tau Robotics' cleaning robots are remotely piloted by humans, collecting training data as they go. The test for me is whether customers rebook at a price that covers the operator.

Switching coding agents shouldn't mean teaching the repo conventions again. Lorekit's call for a shared memory format makes sense to me. I'd want a way to retire a convention across tools too, so it doesn't keep coming back after the code changes.

I'd test a dropped connection halfway through an agent's Blender task. Phaze, now backed by a16z, gives agents access to remote workstations. Can the agent resume from the same state, or does a person have to piece the task back together?

Replaying a user's bug gives a coding agent a concrete test. Replay is proposing automated fixes using deterministic Chromium and Node.js replay. I'm curious how it checks for regressions outside the recorded session.

How much of Anthropic's compute bill is training, and how much is serving users? It rose from 48% to 58% of revenue last quarter. That split would help builders judge whether today's API prices are sustainable.

A same-day vendor replacement also creates a maintenance job. SaaStr says its AI agent replaced a tool during a 24-hour outage. I'd want the handover details: which workflows it covered and who maintains it once the vendor is back.

I'm not sure how a stock drop proves where software budgets went. Sierra added $150M+ ARR this quarter; ServiceNow shares fell 26.9%. Before treating that as a product thesis, I'd want evidence of customers moving spend between them.

An AI answer can feel seamless and still need checking. Apple's John Ternus says users shouldn't have to think about whether AI or traditional software did the work. How do you signal uncertainty without making people learn what's under the hood?