Where do you talk to Amp most: terminal, editor, web, phone, Slack?
80% macOS app, 20% iOS app.
Favorite mode, and when do you switch?
I use Medium (in Fast mode) for the vast majority of tasks, but switch to Ultra when a task is critical, such as designing a new system or double-checking big architectural changes. My opinion is that the vast majority of software engineering tasks (at least as we’ve understood them heretofore) can be done by non-frontier models, and I try to practice what I preach.
Do you still develop locally? If so, when?
No. The only time I’ll “develop locally” is when I need to connect to a local resource that hasn’t yet been integrated into Orbs, such as a non-OAuth MCP server.
What’s a workflow with Orbs that you absolutely love?
As I mentioned earlier, the ability to reference all previous work done at your company (whether yours or your colleagues) via threads is a huge productivity boost. I’d guess that something like 80% of all work is a slightly-different version of previous work, so it’s very nice to be able to tell one agent to reference a previous conversation as context for how to complete a similar task.
How do you review changes?
The first pass is a code smell, which basically reduces to “is this diff bigger than I expected would be necessary when I first asked the agent to work on it”? If yes, I need to dig in — usually the agent has overengineered something, but occasionally there’s something new to learn about our codebase.
From there, I ask the agent to prove that its changes work. You’ll catch a lot of bugs or regressions at this step.
Only once the first two gates have cleared do I actually read the code. I’ll ask the agent to break the diff into logically coherent commits, and then provide a TLDR and file review order for each commit. This allows me to quickly scan for issues while building an understanding of how the diff is actually achieving the outcome I’ve requested.
Which skills, plugins, or MCPs do you use? Anything you built for yourself?
I still use MCP servers a lot in our FDE work, usually to grab data from disparate systems. In our Amp work, we have little need for MCP servers because we’ve integrated everything so tightly.
I find myself using Skills less frequently over time, I think because the most useful and popular skills are quickly RL’d into the model itself.
One meta pattern that I return to, especially when making larger changes that cut across data storage, business logic, and frontend is:
- Use dictation to explain the problem and a few solutions I’ve been thinking about, at length. It is common for me to dictate for 5-10 minutes at this step.
- Have the agent explore the codebase and come back to me with a few implementation options. I can prod, iterate, and refine the plan at this step.
- Ask the agent to implement the plan, pausing during each major milestone for the human review process I described above.
How do you use Puck?
Almost exclusively as a natural language interface for thread search. I search for old threads, whether mine or others, a lot. I think this thread-as-company-knowledge-base pattern is probably the biggest difference between how Amp operates versus our customers.
What’s a prompt or phrase you keep reusing?
Nothing in particular, but two meta points:
- I use dictation a lot, and have for over a year. It’s the best way to get information out of your head and into the agent’s context window. I work pretty closely with our customers, and I’m surprised by the extent to which people are still under-specifying their prompts and then complaining about bad results, especially for ambiguous tasks or out-of-distribution tasks.
- I reference prior threads a lot, as previously described.
What do you still do by hand, and why?
The models exceeded my technical implementation ability a long time ago, but I still make all of the architectural decisions.
There’s very little in our day-to-day work that I do totally au naturel. Writing Slack messages might be the only thing? I always write the final draft of customer comms, but often start with an AI-generated first draft for purposes of efficiency and completeness.
I suppose you could say I’ve become more of an editor than a writer, across both technical and non-technical domains.
Anything else? A surprising thing Amp did, a thing that didn’t work, a thing you wish it did.
If agents were the new thing in 2025, and cloud agents are the thing right now, I think next year will be about agentic CI/CD. We’re already working on it.