Where do you talk to Amp most: terminal, editor, web, phone, Slack?
Mostly web and phone now. I use the web for long-running product and analysis threads, especially when I’m coordinating many Orbs, and the phone for reviewing work, directing follow-ups (a lot on the go: picking up kids, at their various events/practices, running errands, life admin, watching TV, cooking dinner, etc.) and using Puck as an operating layer over everything. Slack is increasingly an actual workflow surface for me! Especially for customer/account intelligence and kicking off investigations (Orbweaver/Legal Puck). I honestly can’t remember the last time I used the terminal.
Favorite mode, and when do you switch?
I tend to lean on high reasoning modes (Ultra) when I’m doing architecture, audits, or ambiguous product work where I want the agent to challenge my assumptions and fill my many many many technical knowledge gaps. I used to use High as my default, but then moved to Medium as the time and outcome weren’t worth it. Medium seems to be the sweet spot. I also follow our generally directed advice on how to use Amp and spin up lots of Orbs with narrow scope and precise acceptance criteria. Medium helps keep me in the flow with this pattern of work the best.
Do you still develop locally? If so, when?
Can’t remember the last time I developed locally. Why? So much friction, especially for a technical dipshit like myself! I love working with Orbs because they isolate work super cleanly and let me run many threads in parallel without tying up my machine or me having to maintain 5+ dev server windows in Ghostty.
I will say that with Orbweaver I started by using a non-durable DB in Orbs to store all the data I want. But, then realized that I needed a more durable DB. So I spun that up and it now replicates the durable DB in my long-running observer thread for testing. That’s been super cool/helpful. Which leads to Amp as Git being mega awesome.
What’s a workflow with Orbs that you absolutely love?
A lead Orb coordinating multiple persistent specialist threads against a large product problem. I can give the lead a broad goal, have it divide the work into non-overlapping investigations or implementation streams, let those threads debate and refine designs with the parent, and then integrate only the work that survives independent verification.
One really specific example is with Orbweaver: Slack “evidence” (Harry is talking about something with a customer in their channel) comes in, the durable services store that info, specialist Orbs (Legal Puck, historical context assignment (it looks for potentially contradictory info like “2 weeks ago you said the workshop was the 9th, now you are saying it’s the 12th” and it flags that discrepancy), and a few more I’m thinking through) investigate it, and agents turn it into attributed, reversible, and logged CRM actions/data.
I also love that when one environment loses credentials or access, a fresh Orb can often independently reproduce the analysis or recover the work. This has saved me so. much. fucking. time. and. headaches.
How do you review changes?
I’m a bit embarrassed to admit this, but I don’t really do any code review. The amount of code I generate and the complexity of the work makes meaningful code review pointless. For me at least.
How I do review breaks down into two buckets:
- For data/analysis work I just look at the output and verify against actuals.
- For Orbweaver stuff, I use portals and require Amp to provide snapshots/videos of the features then test them myself for expected behaviors/outcomes.
I used to ask the oracle to review or use Beyang’s review capabilities, but I found I was even more blindly accepting their suggestions. So, I’ve stopped and opted for a more manual human UI/UX/numbered review.
I find myself bottle-necked doing my review on mobile with Portals. The phone size isn’t the problem, it’s that portals on mobile seem to be very unstable and non-performant.
For backend or analytical work, I have to be at my computer, but I don’t see a realistic world where I do that kind of review on my phone.
Which skills, plugins, or MCPs do you use? Anything you built for yourself?
I built a searching-amp-wiki skill because generic search kept missing our internal operational
information. It refreshes the wiki Git repository into a local cache, searches Markdown, and routes
internal questions — office information, policies, banking, vendors, security — to the wiki before
the public web.
With Orbweaver, I’ve also been building a GDrive integration with search/read tools, plus hosted and self-hosted Drive MCPs.
I don’t really use plug-ins and the skills. The ones I do use are either baked into Amp (I’m honestly not sure which I/the agent use) or are specialized tools where they solve a real workflow rather than accumulating plugins or MCPs for their own sake.
How do you use Puck?
Almost never. I prefer the parent<>sub orb dynamic for orchestration. Building off the above — narrow scope, precise criteria — I’ve found that workflow to be best tied to the orchestrating/parent thread. I also use it for thread search. I find our actual thread search kind of sucks. Puck is much better, especially when I can’t remember the thread name (or, when looking for someone else’s thread(s)) or if I got lazy with sub-threads naming conventions, remember the inter-thread relationships.
I have been using it more for TODOs, setting up cron jobs, etc. recently though.
Maybe I’m not doing enough work that’s “platform wide” so that’s why I don’t find Puck to be useful? I also don’t like having to leave a thread to talk to Puck, especially if I’m in the flow of my work. Have often found myself wishing there was a modal or slash command or something to summon Puck to stay in my thread and parallelize my ask / requested work of him. I know I can just open another tab, but then that’s just another tab I have to manage and it eats up mental capacity. I want something ephemeral or throw away with less of a dedicated interface when in the flow of my work.
What’s a prompt or phrase you keep reusing?
There aren’t many prompts I re-use, but I follow a prompting framework/workflow of something like: investigate first > define acceptance criteria > show reproducible evidence > propose the smallest root-cause fix (when something isn’t right) > then implement > and independently verify it with evidence.
I do find I have to keep prompting agents to use persistent parallel threads rather than disposable subagents when I want them to collaborate, revisit conclusions (“talk with X thread to unfuck it”), to generally orchestrate threads.
What do you still do by hand, and why?
Not much, at least in the context of writing code. The things I want to do far outpace my technical capabilities. The only things I really do by hand are UI/UX validation / clicking and / or manually reviewing the evidence I ask the agent for.
Anything else? A surprising thing Amp did, a thing that didn’t work, a thing you wish it did.
IDK, I think I’m an atypical Amp user. Other tools have native integrations with all the work tools that are super easy to set up and tap into, and a ton more “knowledge worker” capabilities like being able to create Excel / Google Sheet workbooks, presentations, etc. I still have to reach for them a fair amount for that specific work, but I don’t want to! I do find Amp to be a better experience, but for my specific “business guy” needs, there are still gaps.
It’s still fucking awesome how I’ve been able to build so many tools (mainly with Orbs) that can do more and more of my work (Orbweaver CRM, customer knowledge base, customer knowledge base search and reporting, financial auditing and closing, etc.).
I’m about 80% committed to getting rid of my computer and only using an iPad to force me out of using Excel / Google Sheets. I’m excellent with those tools, but I can’t help but think I can do the work I need without them. It’ll be painful, but it will be an interesting test to see.
Orbs + browser-first + portals are so legit. They make the above crazy idea much less preposterous as they make software (now jellyware) immediately tangible and useful for my needs.