Where do you talk to Amp most: terminal, editor, web, phone, Slack?
The Mac app for getting serious work done, and the iOS app fills in the gaps when I’m not at a computer.
I’ll get Amp to DM me on Slack to get my attention when something finishes, but I’m not starting threads there.
Favorite mode, and when do you switch?
Medium with fast mode when I know what I want, Ultra for when I don’t.
Starting this week, I’ve been using High to test out Astra.
Do you still develop locally? If so, when?
I’ll start up an amp --no-tui if I want to give Amp access to local credentials, or need an
urgent answer and don’t want to wait 20 seconds for an orb to start.
How do you review changes?
A lot of my work is security-focused, so trusting the agent or watching a video walkthrough of the agent using the happy path doesn’t cut it.
I usually get the agent to walk me through the problem and the solution to the point where I understand both. Then I usually skim the file change list — I’m not looking at the diff so much as making sure the changed files match my mental model of what I expect to have changed.
For complex changes, I sometimes ask Amp to build a testing harness. I used this recently to test switching the CLI authentication to OAuth. Amp wrote a couple of services to simulate the Amp server and our auth provider, and then spent an hour breaking the new CLI with every conceivable edge case. Every time it found an issue, it fixed it and carried on testing.
Which skills, plugins, or MCPs do you use? Anything you built for yourself?
The most useful skills I use are ones I’ve written for finding and managing security issues in our codebase.
I’ve been iterating on these for a few months, and they’re tuned to work really well on Orbs by fanning out dozens of threads to find and validate issues, each tucked away with their own isolated environment where they can run code, write tests, and harnesses, and then report back to the main thread.
I kicked off a scan this morning that ran for an hour over 40 orbs, and didn’t think about it again until I got a completion notification.
How do you use Puck?
Sparingly. I still haven’t found a good use-case for it.
What’s a prompt or phrase you keep reusing?
If a model has dived deep into an area of the codebase, it happily assumes you also read everything
it did, and that “Nothing extraneous. The friendlyDeviceFromUserAgent Amp-CLI branch is redundant
belt-and-braces with the DB-based labeling […] idempotent, seed is static INSERT ... VALUES, ON CONFLICT (feature) matches the actual PK, SCHEMA.txt updated” is a reasonable thing for it to
say (that’s a real Fable quote).
I’ve found this prompt helps a lot — it’s similar to the “ASD-STE100 Simplified Technical English” prompt that’s going around, but I find it works better for me: