16:35Y Combinator · Interview
Conductor CEO Charlie Holtz Walks Us Through His AI Coding Setup
The founder of an AI agent orchestrator explains how he uses his own product to build his own product and why code is becoming sawdust.
June 4thThe argument in one line.
Giving AI coding agents direct access to the simulator, browser, and telemetry tools they need to verify their own fixes eliminates the manual feedback loop that makes solo development slow.
Read if. Skip if.
READ IF YOU ARE…
- A solo iOS or web developer already using Claude Code or Cursor who still manually pastes logs, screenshots, or error messages back into the chat.
- A developer who wants to walk away from a running agent session and come back to a finished fix rather than babysitting every step.
- Someone paying for AI coding tools and unsure whether Opus max thinking is worth it over Sonnet medium thinking.
- A solo developer shipping many PRs a day with no code reviewer and looking for an automated safety net.
SKIP IF…
- You are new to AI coding tools — the video explicitly skips basics like plan mode, debug mode, and dictation.
- You are not building native iOS or web apps; Xcode Build MCP is iOS-specific.
The full version, fast.
The highest-leverage change in this workflow is closing the verification loop: instead of fixing a bug and then manually checking it, you give Claude Code an iOS simulator via Xcode Build MCP or a Chrome instance via --chrome so it can test its own changes and keep looping until they pass. The same logic applies to production debugging — connecting Sentry, Supabase, and Axiom as CLIs or MCPs lets the agent investigate a crash report in three minutes instead of 45. For code review, Greptile auto-grades every PR and Claude Code is told to iterate until it hits a 5/5. The tool layer runs on cMux (low RAM, sidebar, notifications) and the model setting is Opus 4.7 at max thinking for most tasks, with GPT-5.5 extra-high reserved for complex multi-file bugs where Opus tends to fix one thing and break two others.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:58
01 · Context and model split
Sets up the video as an advanced update. Declares 70/30 split: Opus 4.7 in Claude Code vs GPT-5.5 in Cursor.
00:58 – 02:00
02 · Automated testing — iOS (Xcode Build MCP)
Xcode Build MCP lets Claude Code control the iOS simulator: build, tap, screenshot, read logs. Keeps Xcode closed 90% of the time.
02:00 – 04:02
03 · Automated testing — web (claude --chrome)
claude --chrome gives Claude a live Chrome instance. Cursor has a built-in browser. Old manual loop diagram vs. new autonomous loop.
04:02 – 06:26
04 · MCPs and CLIs for production debugging
Sentry CLI, Supabase MCP, Axiom CLI all wired into Claude Code. Crash investigation drops from 45 minutes to 3. CLI preferred over MCP for token efficiency.
06:26 – 08:24
05 · Greptile for automated code review
Sponsored. Every PR auto-reviewed and scored 1-5. Claude Code loops until 5/5. Beat Cursor BugBot across 60 real PRs.
08:24 – 10:01
06 · Remote control — code from your phone
/remote-control syncs session to the Claude mobile app. Boris (Claude Code creator) called this his top hack. Config tip: auto-start on launch.
10:01 – 11:23
07 · cMux as terminal
Project sidebar, per-instance notifications, runs 20 Claude Code instances on 64GB M4 Max without issue vs. 5 instances choking in Cursor.
11:23 – 13:33
08 · Claude Code settings: model and effort
Opus 4.7 + max thinking on $200/mo plan. Flag to auto-start max thinking. Most people complaining about quality are on Sonnet medium.
12:53 – 13:33
09 · No-flicker mode
CLAUDE_CODE_NO_FLICKER=1 beta flag: pinned input bar, clickable cursor position.
13:33 – 15:15
10 · When to switch to Cursor + GPT-5.5
Complex multi-file bugs where Opus fixes one thing and breaks two others. GPT-5.5 extra-high at 1M context wins here. Burns Cursor Ultra credits fast.
15:15 – 15:38
11 · Why not Codex
Has the $100/mo plan but prefers Cursor UX. Will revisit if Codex iOS app ships.
15:38 – 17:06
12 · Recap — what to steal
Four-point summary: automated testing, MCPs/CLIs, Greptile, slash config. Ends with dog cameo.
Lines worth screenshotting.
- Giving Claude Code a running iOS simulator via Xcode Build MCP eliminates the need to open Xcode at all for 90% of development tasks.
- Running claude --chrome hands Claude Code a live browser so it can test, screenshot, and read the console without any copy-pasting from the developer.
- Connecting Sentry, Supabase, and Axiom as CLIs or MCPs compresses a 45-minute manual crash investigation into a 3-minute autonomous one.
- CLIs are preferable to MCPs when both exist for the same service because they consume fewer tokens and agents navigate them more reliably.
- Telling Claude Code to wait for Greptile to finish reviewing, fix every real comment, and loop until it scores 5/5 automates the entire post-PR QA cycle.
- Over half the developers who complain Claude Code is not smart enough are running Sonnet at medium thinking, not Opus at max thinking.
- cMux runs 20 parallel Claude Code instances on a 64GB M4 Max without issue; the same machine struggles with 5 instances inside Cursor.
- CLAUDE_CODE_NO_FLICKER=1 pins the input bar to the bottom and makes the cursor position clickable.
- GPT-5.5 extra-high with a 1M context window outperforms Opus max thinking on bugs that touch 50+ files across multiple repos.
- At $400/month in AI tooling, the breakeven is whether the time saved building apps exceeds the cost — for a full-time indie developer, it does.
- Remote control via /remote-control was the top workflow hack recommended by the creator of Claude Code himself.
- Setting a config flag to auto-start remote control on every session means continuity from desktop to phone is always on.
- The Greptile vs. BugBot comparison ran across 60 real PRs over six weeks — not a synthetic benchmark.
Close the loop before you walk away.
WHAT TO LEARN
The reason AI coding agents stall is not that they are not smart enough — it is that they cannot see whether their fix worked, so they stop and wait for a human who is not there.
- Xcode Build MCP gives an AI agent the ability to build an iOS app, run the simulator, tap the UI, take screenshots, and read crash logs without the developer opening Xcode or pasting anything.
- Running claude --chrome hands the agent a live browser session so it can test web changes, read console errors, and iterate without any manual log transfer.
- Connecting Sentry, Supabase, and backend log tools as CLIs or MCPs compresses a multi-service crash investigation from 45 minutes of manual tab-switching to a 3-minute autonomous query.
- When both a CLI and an MCP exist for the same service, the CLI is preferable: it uses fewer context window tokens and agents navigate it more reliably.
- Automated PR review closes the quality loop the same way: tell the agent to open a PR, wait for the review score, fix every real comment, and loop until it passes.
- Most complaints that an AI coding tool is not smart enough turn out to be a model and effort setting problem: Sonnet at medium thinking and Opus at max thinking are not comparable tools.
- For bugs that touch 50+ files across multiple repos with a dozen interacting edge cases, a model with a 1M context window is structurally better suited than one with a smaller window.
- Running 10-20 parallel AI coding sessions requires a terminal built for that use case — one with per-session notifications and a low memory footprint — not a general-purpose IDE terminal.
- Remote session continuity matters most when a session takes 20-30 minutes and the developer cannot stay at their desk for the duration.
- The /config settings in Claude Code are not defaults worth accepting: model choice, thinking level, remote control auto-start, and rendering mode all have better values than what ships out of the box.
Terms worth knowing.
- Xcode Build MCP
- A free MCP server made by Sentry that gives Claude Code the ability to build iOS apps, run the iOS simulator, take screenshots, tap UI elements, and read logs without Xcode being open.
- cMux
- A terminal application for running multiple Claude Code sessions simultaneously, with a project sidebar, per-instance notifications, and a much lower memory footprint than running Claude Code inside Cursor.
- No-flicker mode
- A beta rendering mode for Claude Code (enabled via CLAUDE_CODE_NO_FLICKER=1) that pins the input bar to the bottom of the terminal and makes the text cursor position clickable.
- Greptile
- An AI-powered code review service that automatically reviews pull requests, surfaces issues, and assigns a 1-5 quality score when connected to a GitHub repository.
- Max thinking
- The highest reasoning effort setting available in Claude Code. The default when starting a new session is medium thinking, not max.
- Remote control
- A Claude Code feature (/remote-control) that syncs an active terminal session to the Claude mobile app, letting the developer continue or monitor a running session from their phone.
- Axiom
- A log management and analytics platform used here as a backend logging service, accessible via CLI so Claude Code can query server logs during debugging.
Things they pointed at.
01:38toolXcode Build MCP
03:00toolclaude --chrome
10:01toolcMux
Lines you could clip.
03:45
“I let Claude code run for like twenty, thirty minutes, and then by the time I get back, the issue is fixed.”
04:50
“The old me would spend an hour digging through logs... now Claude Code does it in like three minutes.”
12:13
“Over half the time they have it set to SONNET and medium thinking. And I tell them, okay, well, that's why.”
13:45
“GPT-5.5 extra high, 1,000,000 context window is better than Opus 4.7 max thinking at majority of tasks, especially if the task is very complex.”
15:05
“$400 a month in AI tooling, but because I make a living building apps, that cost is totally worth it.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogystory
00:00It's been a few months since I've talked about my coding workflow, but so much has changed that I figured it was time for an update. In this video, we're gonna go over all the new stuff I've added to my coding workflow, why I'm paying for both Cloud Code and Cursor, how I'm doing automated testing on both iOS and web, and a few features that have completely changed how I work.
00:16If you're new here, welcome to the video. My name is Chris, and I build productivity apps. And today, we're gonna be going over my coding workflow.
00:21A quick heads up, this is more of an advanced workflow video, and I really wanted to focus on what's new with my workflow. If you want the basics on how I use plan mode, debug mode, how I use dictation, I will leave a link in the description to my other videos. Today is about the newer, less obvious stuff.
00:35Real quick, let's talk about the models that I'm using before we dive in. I'm using Opus 4.7 in Claude code, and specifically, I run everything in max mode.
00:43And I'm using GPT 5.5 extra high in cursor. Both of these are the latest models at the time of recording. 70% of my coding is in Claude code, and the other 30% is in cursor.
00:52And I'll cover when and how I switch later in the video, but first, let's go over some of the cool new stuff that I wanna show you guys. The biggest unlock I've had over the last few months is getting Claude Code and Cursor to use my app and help me automate testing. Usually, way that I, and I'm pretty sure a lot of you guys test, is you have Claude Code implement a fix, and then you have to run the simulator and then check and make sure everything works.
01:13On iOS specifically, I found something new called Xcode Build MCP. It's a free MCP tool made by the people at Sentry, and it lets Cloud Code and Cursor basically do 90% of what you can do in Xcode. When you give Cloud Code this MCP, it can build the app, it can run the simulator, it can take screenshots, it can tap on things on the screen, and it can pull all of the logs.
01:33So now whenever I'm fixing a bug, I just have to tell Claude Code, can you make this change? And then verify it works in the simulator using Xcode Build MCP. And when it's done, if everything's working properly, you should just see the simulator open, and ClaudeCode will take control of the simulator and start tapping and testing for you.
01:49And it has the ability to take screenshots and access the logs, so no more copy and pasting logs from Xcode into ClaudeCode or dragging screenshots into ClaudeCode. It is all handled for you. And because of this, I can actually keep Xcode completely closed when I'm building iOS apps.
02:03I can just have Claude Code open and it'll do everything Xcode can do. There are still a few things that Xcode Build MCP can't do. It's like maybe 10% of things, and that includes changing Xcode specific settings where it's easier to just open up Xcode and change it directly or testing anything that requires complex multi gestures.
02:19So in my daily planning app, LE, I have this ability to reorder a list. Unfortunately, Xcode Build MCP does not have the ability to test this for me, so I have to manually do that part myself. But it does have access to the log, so once I'm done manually testing that myself, I can have it read the logs directly, and at least that part is automated.
02:36But outside of the multi touch gesture stuff, 90 of everything I need is handled by this MCP, and, again, I just don't have Xcode open most of the time. On the web, it's actually much easier. For Claude code, I'm specifically using something called Claude with Chrome.
02:49It's kind of the same idea, but instead of giving Claude code a simulator, you can actually give it an instance of Google Chrome. And the way you run this is first you install the Claude Chrome plugin, and then you run Claude dash dash Chrome, and then you'll see that Claude has the ability to control Google Chrome.
03:03So now I can test my app, take screenshots, and even read the console. So no more pasting logs between Chrome and Cloud Code anymore. And if you're using Cursor, it's even easier because they have a built in browser that does all of this stuff for you.
03:14Previously, my workflow was I told Cloud Code something's broken, it goes and fixes it, then I manually have to go verify that the thing is fixed, and then if it's not, I have to take screenshots, I have to paste console logs, and then I have to go back to Cloud Code and say, hey. It's still broken. Can you try fixing it again?
03:29And then we repeat this process over and over again until the issue is fixed. But now with the Xcode Build MCP and Cloud with Chrome, I just have to tell it, hey. Can you fix the problem?
03:38And then go verify it by using Xcode Build MCP or Claude with Chrome and keep going until you fix it. And I let Claude code run for, like, twenty, thirty minutes, and then by the time I get back, the issue is fixed. This is game changing because it makes context switching much easier because I can let ClaudeCode run for a longer period of time without intervention.
03:56While we're talking about automating things, the second biggest unlock I found is using MCPs and CLIs way more than I was a couple months ago. When I made my last Claude code workflow video, I was barely using MCPs.
04:07I think I was just using the Supabase MCP. But now I am fully addicted to using MCPs and CLIs, and I genuinely cannot imagine working without them now. Let me show you what this looks like in practice with a production application of mine.
04:19If a user emails me and says, hey, Chris, the app keeps crashing when I open the settings page, the old me would spend an hour digging through logs. I would have to go into Sentry, which is what I'm using to look at crashes in the apps. I would have to go into my database, look at the user's data to figure out is there something that got corrupted, is there something weird going on.
04:38And then sometimes I would have to go into Acxiom to go look at server logs. I would have to go to a bunch of different services to piece together what happened at the time of crash. But now because I've hooked up all of these services as MCPs or CLIs, I can just ask Claude Code, hey.
04:53This specific user, here's their email. They reported a crash in the settings page. Can you go figure out what happened?
04:59Then And I tell it, feel free to check out Sentry, Acxiom, or Supabase, or Firebase using the CLIs or MCPs you have available to debug this. And now what it's going to do is it's going to use the Supabase MCP to pull up that user's account info. It's gonna check if the settings look weird.
05:13It's gonna pull Sentry to see if there's any crashes for this specific user. And within Sentry, it's also gonna check what specific iOS version is the user using. Is there anything weird that sticks out with their specific device?
05:24Like, they on an older operating system or something? And if you configure Sentry correctly like I did, it'll even show you where in the code the crash actually occurred. And if for some reason it needs the back end logs, it has the ability to pull Acxiom, which is what I'm using for all of the logging as well.
05:37Hooking up these tools directly into Cloudco and Cursor has been game changing for my workflow because now I don't have to go in to all of these services and try to piece everything together, which would have taken me, like, forty five minutes. Instead, I can just have Cloudco do it, and it will figure all of this out in, like, three minutes.
05:53Almost every major developer service has some sort of MCP or CLI available, and there's really no reason not to use it. Now quick tip, if you're trying to figure out should I use the MCP or the CLI, my default is to actually use the CLI whenever possible. The reason is after talking to some engineers who know way more than I do, they told me that in most cases, the CLI is a bit more efficient in terms of the context window, so it eats a lot less tokens than the MCP servers.
06:17And in general, I feel like the agents just know how to use the CLIs a little bit better than the MCPs. But honestly, you can't go wrong. If you use either, you should be good.
06:25Next thing I wanna highlight in my workflow is what am I using to do code reviews? I get this question a lot on my channel, and I have recently started using a service called Reptile, and a huge shout out to them for actually sponsoring this video.
06:36But even if they weren't sponsoring, this is the tool that I would be using anyway. Greptile is an AI code review service, so every time I open a pull request on any of my apps, Greptile will review the pull request and try to service issues, which is huge for me because as a solo developer, I don't really have anyone reviewing my code, and I'm shipping a ton of PRs every single day.
06:55I was previously using Cursor's BugBot, but I've switched over to Greptile, and I don't know how they pulled this off, but it is really good at reviewing code. I ran BugBot and Greptile simultaneously on, like, 60 p r's over the last month and a half, and Greptile has been noticeably better.
07:11If you want proof of what this looks like in practice, I recently open sourced my AI agent so you can go check out the code, but I have Greptile set up on the repo. So if you go to the pull request tab, you can see Greptile running and reviewing each of the pull requests, and you can see Greptile in action just by looking at the comments.
07:27I'm pretty sure on, like, 95% of the pull requests I've opened, it has caught at least one valid thing. The setup is dead simple.
07:33You sign up for Greptile, you link the repos you want it enabled on, you toggle it on, and then from that point, every PR that you open will get automatically reviewed. And here's my specific workflow, which has been a huge unlock for me. When I'm shipping a feature, I open a pull request, and then Greptile's gonna review it.
07:48Then I tell Claude Code, can you wait for Greptile to finish reviewing? And then go through every comment, fix anything that's real, and keep looping until Greptile gives you a five out of five. Reptile actually grades your PR with a score, and so I basically have Claude code just looping and iterating until it hits a five out of five.
08:05And that's when I know that the PR is ready to ship. I'll leave a link in the description if you wanna check Reptile out. You get fourteen days to try it.
08:11No credit card required. I think the service is $30 a month, and it is well worth it for the peace of mind that it gives me, especially as a solo developer who has no one to review the pull request for them. That's the newest tool that I've added to my workflow, and I've been seriously impressed by what it's been able to catch.
08:25The next feature I wanna talk about is something that I thought everybody knew about, but people have been pretty surprised when I bring it up, and it's called remote control. Basically, it's a way for you to continue a session that you have running on Cloud Code on your phone or desktop, but I primarily am using it on my phone, which is huge for me because I'm constantly having to leave the house and run errands every time I'm in the middle of some sort of coding session with Claude Code.
08:46The way it works is if you have Claude Code running on your computer, you can type slash remote control, and it lets you continue the exact same session on your phone. So if you open the Claude app, you'll go to code, you'll see that session appear there, and you can continue it like nothing happened, and it will sync between the two devices.
09:03Again, huge if I'm in the middle of a coding session and then I have to to the grocery for some reason, I can then continue this session while I'm on the go. Probably a little unhealthy, but I will own up to that. I actually had the opportunity to talk to Boris himself, who's the one who created Cloud Code, and I asked him, what is the coolest feature and the biggest workflow hack that you have?
09:20And this is the one that he mentioned. He also gave me another tip that you can actually run slash config, and then you can have it set where every time you start Cloud Code, it's going to automatically start a remote control session.
09:31So now I have this running as well. You're probably wondering, is there something for Cursor? There are Cursor Cloud Agents, and they work really great.
09:38I still use them for certain things. But the reason I prefer Cloud Code's remote control is because Cloud Code has access to all of the MCPs and CLIs I have on my device. But when you're using Cursor's Cloud Agents, you really only have access to a handful of remote MCPs.
09:53Now there are some workarounds to get specific CLIs and some other local only MCPs working, but it is such a hassle to set up, and I've just been too lazy to do it, to be honest. Now in terms of what I'm using to run Cloud Code, I'm actually using an app called c mux. And then Cursor is obviously just Cursor.
10:07But if saw my other videos, I used to run Claude code directly in the Cursor terminal. I have recently switched over to c mux, and here are a few reasons why. The first is that c mux has a really nice sidebar, so I'm able to see all of my projects on the left, and then I can open up as many windows as I want on the right.
10:23And it's really intuitive to do this, but it makes juggling multiple project, which I'm constantly doing, really easy. The second reason is it has built in notifications. So when a specific Cloud Code instance needs my attention, cMux will ping me, and then when I click it, it'll automatically open to the relevant window, and it actually even highlights which window needs my attention.
10:41And this is really important if you're like me and you're running like 10 Cloud Code instances at once at all times. Now the third reason and arguably the most important reason I'm using cMux is it is incredibly lightweight when it comes to memory on my device.
10:55I've noticed that when I'm using Cloud Code inside of Cursor, this thing is consuming RAM, and my computer gets really slow.
11:03And I have a very, very powerful computer. I got, like, 64 gigs of RAM, and this is an m four max. But it still struggles when I have, like, five instances of Clodder code running inside of Cursor.
11:15I rarely have that problem when I'm running, like, 20 instances of Cloud Code inside of CMUX. It does just fine. The memory footprint is really light.
11:23Now in terms of Cloud Code itself, the model I'm using is Opus 4.7, and I'm specifically running it on max mode, which is the highest thinking mode that you can set it at. I'm on the $200 a month plan, and I've had zero issues with the rate limit so far.
11:36Another tip, whenever you start a new Cloud Code session, it is going to default to medium thinking. I got kind of annoyed having to set it to max every single time, and I realized there's actually a flag that you can set. And when you set this, every new session is automatically gonna start with max thinking enabled.
11:52So I have that set right now. Honestly though, you could get away with extra high or even medium in most cases, but I set it to max just because I have the $200 a month plan. On that note, a lot of people keep telling me, I feel like Claude Co.
12:03Just doesn't get it. It's just not that smart. And in most cases, I usually tell them, can you check what model you're using and what thinking mode?
12:09And over half the time, they have it set to SONNET and medium thinking. And I tell them, okay. Well, that's why.
12:15That is not that great of a settings. I mean, it's probably good for most coding tasks, but in my case, almost all the tasks that I run, I have to run them in Opus and Max mode to get the best results. Another setting I have turned on is that I'm using Claude code in what's called no flicker mode.
12:29So I think this is a beta thing at the time of recording, but they have this new rendering mode where it'll render a little differently in the terminal compared to if you have this off. You don't lose anything out in terms of functionality, but there are a few key differences. Like, for example, when I scroll through a conversation, the text input bar stays pinned to the bottom.
12:48There's this little button that appears where I can click it to scroll to the bottom. It's also just a little bit smoother when it comes scrolling too. But the biggest thing that it gives you is you can click directly in the text input field, and the cursor will jump to that position.
13:01It's really annoying on the default Clawdocode experience where you have to keep tabbing to move your cursor because for some reason clicking does nothing. There's a few other small quality of life improvements that come when you use no flicker mode, but these are the two I wanted to highlight. You could probably go online and check the full list, and then I would just enable this, and it does feel like a meaningful step up in terms of experience.
13:19So this is the flag you need to set when you do it. No flicker mode will be turned on. I mentioned at the beginning of the video that I'm also using cursor for about 30% of my tasks, so why am I doing that?
13:28Cloud code seems completely fine. Well, are some cases where Opus 4.7, even with max thinking, does get the task.
13:36And specifically, this happens with very complex bugs with a lot of edge cases. For those bugs, that's when I switch to cursor, and I specifically use GPT 5.5 extra high with the 1,000,000 context window.
13:48Probably the most controversial thing I'll say in the video, but in my experience, GPT 5.5 extra high, 1,000,000 context window is better than Opus 4.7 max thinking at majority of tasks, especially if the task is very complex.
14:03I'll give you an example. In my daily planning app, I have this feature called recurring tasks. It is a very complicated feature.
14:09It touches like 50 parts of the code base across multiple repos. It's a very complex system. And when something breaks in the system, I have to consider like a dozen different edge cases.
14:18Every time I use Opus 4.7 for this, it fixes the bug, but then two or three other things break in the process. But I found that GPT 5.5 with extra high usually is able to think through these edge cases and solve it in one or two goes. You're probably wondering, why don't I just use GPT 5.5 extra high for everything then?
14:36The answer is cost. It is a ridiculously expensive model to run.
14:40I'm on the $200 a month Cursor Ultra plan, and I've noticed that all the credits get evaporated after a few days of using this model. But it's also a massive overkill for, like, 90% of the tasks that I'm working on. So my rule of thumb is I try to use Opus 4.7 and Cloud Code for the task, but if it's struggling to get it or I have a hypothesis that it's not going to get it, then I switch to GPT 5.5 extra high with the 1,000,000 context window.
15:03And yes, I have the $200 a month max plan and the $200 a month ultra cursor plan. So that's $400 a month in AI tooling, but because I make a living building apps, that cost is totally worth it, to be honest. I'll probably get a couple comments asking, why don't I just use Codex instead of Cursor?
15:19I do actually have the $100 a month Codex plan just to test things out, but in my experience, I just enjoy the Cursor experience a lot more where I prefer to use it whenever possible. But I will say Codex has an extremely good Mac app and I heard there's rumors that they're gonna release an iOS app, so I'll be constantly testing it and if anything changes, you guys will be the first to know.
15:37So a quick recap here, here are the things that I would steal. The first is use Xcode Build MCP if you're building an iOS app, you don't have to use Xcode anymore. And if you're building on web, use Claude with Chrome or Cursor's web browser to do your testing.
15:50If you have not tried to automate the testing workflow, please do it because it's one of those things that doesn't click until you do it. But now that I've done it, I actually just cannot go back. Second, try to use MCPs and CLIs wherever possible.
16:01It has been game changing in terms of production debugging. If you haven't done it before, it is also game changing. I feel like I've been using that word a 100 times in this video, but really, these things are actually kinda game changing for my workflow.
16:12Third, get some sort of AI code review system in place like Reptile. It is massive for me as a solo developer. I don't have anyone to review my p r's and it really helps me sleep at night.
16:21Last, spend some time looking at the Claude code configurations. If you type slash config, you should see a list of settings.
16:27Go through all of them and really customize Claude code to work with your workflow. And if you haven't tried remote control and coded on your phone before, please do it. I will warn you, it is kind of addicting and probably a little unhealthy, but it has been massive in terms of being able to continue sessions while I'm on the go and just not disrupting my workflow.
16:43That's basically all the new stuff. If you want the basics of my workflow, again, I will leave a link in the description to my two videos on ClaudeCode and Cursor. But if you like this content, check out my Instagram and TikTok.
16:53I post almost every other day about building productivity apps. And, obviously, if you like this content, don't forget to subscribe. But thank you guys so much for watching, and I will see you guys in the next video.
The Hook
The bait, then the rug-pull.
Chris Raroque builds productivity apps solo and ships PRs every day — and for months, every fix meant a manual loop: implement, switch to Xcode, run simulator, check, screenshot, paste back, repeat. This video is about what replaced that loop: AI agents that test their own work, review their own PRs, and keep iterating until the job is done.
Named ideas worth stealing.
03:00model
The Autonomous Verification Loop
- Tell agent to fix the problem
- Tell agent to verify using Xcode Build MCP or claude --chrome
- Agent loops until verified
- Developer returns to finished fix
Instead of the developer verifying each agent fix manually, the agent is given the tools to verify its own work and instructed to keep looping until it passes.
Steal forAny agentic coding workflow where the bottleneck is the human checking if the fix worked
08:03model
The Greptile PR Loop
- Open PR
- Greptile auto-reviews and scores
- Tell Claude Code to fix all real comments
- Loop until Greptile gives 5/5
- Merge
Replaces human code review for a solo developer by chaining AI code review with AI remediation until a quality threshold is met.
Steal forSolo developers shipping fast who have no one to review their PRs
05:50concept
CLI over MCP
When a service offers both a CLI and an MCP, prefer the CLI: it uses fewer context window tokens and agents navigate it more reliably.
Steal forDeciding how to wire developer services into Claude Code or Cursor
How they asked for the click.
VERBAL ASK
16:18subscribe
“if you like this content, check out my Instagram and TikTok. I post almost every other day about building productivity apps. And obviously if you like this content, don't forget to subscribe.”
Soft and personal. Directs to social channels before subscribe ask. No aggressive pitch.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
