Where do you talk to Amp most: terminal, editor, web, phone, Slack?
80% MacOS app, 10% Slack, 5% phone, 5% terminal.
I mostly use Slack to kick off threads to easily use the context from that Slack conversation, then iterate and review in the MacOS app. Phone is great for monitoring things, but I still hate typing on my phone and I feel awkward about talking to my phone in public, so the “real” work I get done is still mostly on the MacOS app. And terminal is really only for the occasional laptop setup question or rarely when I want to use my personal accounts that I haven’t (or don’t want to) set up in orbs. Mostly high-risk stuff.
Favorite mode, and when do you switch?
Ultra. Fable 5 was good but too expensive to make it feel worth it. Fable 5.1 in Ultra just knocks it out of the park. It’s amazing how much faster I can move when I can start to trust delegating more tasks autonomously. The other modes are still good and I still use them for quick questions or for grunt work or for changes which are already well-scoped, but anything which requires any sort of architecture decisions pulls me back to Ultra.
Do you still develop locally? If so, when?
No.
What’s a workflow with Orbs that you absolutely love?
Simple, effective, and often also exposes some metrics or signals that we should add to our regular dashboards and/or alerts.
How do you review changes?
Completely depends on the task.
I almost always do an architecture review, but mostly by asking probing questions rather than reading the code. Looking for things like whether the change will be a backcompat nightmare, or whether it exposes more surface area for security vulnerabilities.
I still skim the code, mostly just to confirm it matches the shape of change I expect. Occasionally, Amp still slips in things like backcompat shims for unreleased features, or unnecessary caching, or fails to clean up code that was made obsolete by its change, and those things add up over time and make it harder for Amp to operate smoothly in the codebase.
And the best tests are end-to-end, so I usually ask Amp to test its work against the dev environment. Or, for infra work, deploy it to staging first and ensure there’s nothing unexpected.
Which skills, plugins, or MCPs do you use? Anything you built for yourself?
No MCPs. No plugins except the occasional “try out a new model” plugin. I use a handful of skills, but they’re mostly just named sets of instructions for tasks I do regularly. Things like updating dashboards or monitoring deployments.
I do not use any skill bundles or curated workflows. I don’t find them necessary, and I think they often just increase the number of abstraction layers between me and the model. Over-constraining the model leads to worse results in my experience, so I prefer to keep the “how” mostly unspecified.
How do you use Puck?
Mostly just through Slack. Occasionally as a fancy thread search. Occasionally for super quick general questions that don’t need an orb. No fancy thread orchestration workflows. I prefer to keep my thread relationships simple.
What’s a prompt or phrase you keep reusing?
Just feels right to be polite, ya know? Maybe that’s just the midwest in me leaking.
What do you still do by hand, and why?
Tasks which have a risk of data loss. I do still use Amp to help with these, but I always dictate exactly the procedure I want to use, and I apply the changes by hand so I can validate each step of the way and react accordingly if things go wrong.
Anything else? A surprising thing Amp did, a thing that didn’t work, a thing you wish it did.
I really wish Amp were better at interpreting graphs + timeseries. It’s surprisingly bad at identifying patterns which are obvious to humans, and consistently miscategorizes the urgency of various signals. I think that’s the biggest thing blocking me from leaning on it more heavily for automated monitoring and triage.