Using Amp

How I use Amp

Hitesh Sagtani

Engineer · @sagtanih

Last updated

Where do you talk to Amp most: terminal, editor, web, phone, Slack?

I am mostly on web instead of phone since I am at my desk most of the time. My usage of terminal briefly went up mainly for building HTTP orbs (and was because of dev tunnel via ngrok) but I eliminated that with Tailscale. Also when building Modal/Vercel my creds were in computer so I used terminal for that, think I felt a bit lazy setting up orb environment just for that.

In the last weekend I had to travel to a different city for some personal work and I started something on the laptop and it was nice iterating on the phone with it. Although review was a bottleneck for me and I couldn’t get that feature shipped because there were a lot of changes and review seemed like a blocker (maybe I should get a tablet for such case). But for following up with threads I used phone pretty often when travelling.

Favorite mode, and when do you switch?

I now kind of use Fable all the time. I think now that we are working in async mode where you kick off a bunch of threads and come back to a working solution is so helpful, and Fable at this point seems like one-shot everything.

Do you still develop locally? If so, when?

Just a case when iterating with Modal/Vercel recently because all my creds were in local computer and felt lazy to setup. I think Puck is a good use case here: “Can you sync my MODAL_API_TOKEN to amp workspace”.

How do you review changes?

I think it’s asking for a proof of work in terms of video/image and relevant log lines and then portals to play with. During debugging I also often ask it for a timeline of events from the logs. For simple changes and UI fixes I don’t look at code that much in detail and for core changes I still go through the diff changes.

Which skills, plugins, or MCPs do you use? Anything you built for yourself?

agent-browser is the most used I think in orbs. Built a couple for debugging Substrate and they are mostly backed by scripts, like collect-substrate-logs, load-test-cluster, simulate-10min-amp-workspace-analyze. I think with these the agent was able to check the lifecycle of a sandbox: did it resume on time, error on resume, did it respect the 5 min pause, any errors during peak load, did it trigger a cold node bootup or were warm workers used.

How do you use Puck?

One main use case is finding previous archived threads and reviving them, I use this pattern a lot. Recently: “Tell me my open bug reports”, “Is someone working on feature X?”

What’s a prompt or phrase you keep reusing?

I think with Fable my prompts are very short, I give it minimal direction and it just works with proof.

What do you still do by hand, and why?

pnpm fix sometimes. If I press the Ship button it almost takes about 10-20 min to get to main, it’s a bit annoying; a lot of times pnpm test is on the full repo and it’s so slow.