Time Capsules
Recordings of Thorsten working with Amp, from the Chronicle.
Where do you talk to Amp most: terminal, editor, web, phone, Slack?
On the web and in the macOS app. Then on my phone, using the iOS app. Then Slack.
Favorite mode, and when do you switch?
Right now using High a lot to try out GPT-6 Astra. In the past few weeks I’ve also started to use Ultra with Fable a lot more, because I’ve done a lot of things that involve UI and design and I think Fable has been consistently better with that.
Do you still develop locally? If so, when?
Develop? No. Not in the last 8 weeks. I run Amp locally occasionally to help me with some things
like “Upload all of these files to this thing”. And I run amp --no-tui on my Raspberry Pi for
home automation (it’s incredible at that). But it’s been a long time since I’ve actually ran Amp
locally to write code for me.
What’s a workflow with Orbs that you absolutely love?
One thing that I’ve come to really appreciate is the ability for orbs to spawn up other orbs. It sounds like novelty feature at first, but it’s incredible how much it allows you to automate.
To give you an example: right now (I’m looking at my sidebar here) I’m optimizing a specific part of our codebase to be more agent-friendly, I want agents to check performance in production via a canary before they ship something that touches a certain database. So I have one orb that writes the scripts and tooling and makes changes to AGENTS.md files and every time it makes a change, it pushes them, spawns 3 new orbs and tells them to do a certain task (without mentioning the change we made). Then it waits for those orbs to report back and once they’re done it reads their threads (also super handy) and compares how they did. Then it tweaks, pushes, tries again. Orbs as evals for how agent-friendly the codebase is.
How do you review changes?
First, on a functional level: I ask the agent to test everything, run through everything end to end, and give me “irrefutable proof that it works.” For very important parts of the codebase, I read/scan the code and might even push back on something with “No, this shouldn’t call that”. But I don’t read all the code and for a lot of changes I don’t read it at all.
I think feature flags play a bigger role here. Agents in our codebase are incredibly good at using our feature flag system. So I often have the agent build something, tell it to put it behind a feature flag, roll it out to prod, enable the feature flag for me (they have a tool via an Amp plugin to do that) and then I test it. They are really good at isolating changes with flags. So then once I test it, I turn it on for us internally, and then we smoke out other problems.
Which skills, plugins, or MCPs do you use? Anything you built for yourself?
I don’t know, to be honest. We have skills in our repository, but I’ll need to check what they are. Okay, just checked. We do have a lot of skills for feedback loops and how to use certain things: how to connect to gcloud, how to use agent-browser, how to use BigQuery, how to build the native apps, etc. I use those. But I don’t have any personal skills I download from somewhere and use.
How do you use Puck?
I use Puck as an orchestrator and ask it to manage my threads. For example: I had two threads with automations set up, but not in the right way. Told Puck to recreate them, with better instructions, but with mode set to High (which is Astra). I also use it to talk to on the phone and create new threads. And via the Slack integration, obviously.
What’s a prompt or phrase you keep reusing?
I got that from Quinn and it’s been a game changer.
What do you still do by hand, and why?
Writing prose. I’m allergic to LLM-written prose and I think that goes for everyone on the team. We also all still write Slack messages by hand. We had some automations that would auto-post to Slack and they got quickly killed.
Anything else? A surprising thing Amp did, a thing that didn’t work, a thing you wish it did.
One thing that I love and that I think might be unique to Amp and our codebase: our agents can fully roll out a change to production and monitor it and it’s quite magical to see. You ask an agent in an orb to ship something, it pushes, it checks CI, it waits for it to be deployed, it monitors production traffic. It sets itself up with automations and checks every X minutes and reports status back. It’s incredible to watch.