Big Idea

The argument in one line.

Entering plan mode without first having a detailed conversation gives the agent too little context to plan well -- the conversation IS the planning, and plan mode is just the commit step.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Cursor, Claude Code, or another agentic coding tool and find the agent regularly builds something close-but-wrong on the first pass.
You reach for plan mode as your first move and wonder why the resulting plan is full of assumptions you didn't intend.
You work across multiple agent tabs simultaneously and need a system for keeping parallel workstreams coherent.
You're building a product with a small team or solo and want the agent to behave more like a collaborator than a code-executor.

SKIP IF…

You already have a disciplined prompt-engineering workflow and are looking for framework-level architecture content -- this is workflow, not theory.
You're not using agentic coding tools at all; the advice is specific to Cursor/Claude-style multi-turn agent sessions.

TL;DR

The full version, fast.

Most people use plan mode by typing a single instruction and letting the agent plan immediately -- which means the agent fills the gaps with assumptions, producing a plan that doesn't match what you actually had in mind. The alternative is to spend real time in conversation first: discuss requirements, surface tradeoffs, agree on constraints, and only trigger plan mode once the agent clearly understands the target. After plan mode runs, save the output as a markdown file so both you and future agent tabs have a shared reference. The demo builds a mobile companion app for Pluto, choosing Expo over native SwiftUI after a structured back-and-forth about cross-platform tradeoffs.

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 →

Chapters

Where the time goes.

00:00 – 01:02

01 · Hook and premise

States the claim: his planning method beats standard plan-mode use. Frames plan mode as a commit step, not an entry point.

01:02 – 03:15

02 · How most people use plan mode

Live demo: single-line prompt into plan mode in Cursor. Argues the agent is making too many assumptions when it has one line of context.

03:15 – 05:10

03 · The method: conversation first

Explains his workflow: multiple back-and-forth exchanges, discussing tradeoffs, confirming the agent understands intent, then entering plan mode.

05:10 – 06:05

04 · Why save the plan as a markdown file

The .md file is for the human as much as the agent -- it lets parallel tabs share context and lets you return to a feature days later without re-explaining.

06:05 – 07:00

05 · Sponsor: Depot CI

Sponsor read for Depot CI -- a faster CI engine positioned as an alternative to GitHub Actions.

07:00 – 11:40

06 · Live demo: planning the Pluto companion app

Full real-time demo of the conversational planning method. Prompts the agent to brainstorm companion app ideas, reviews output, discusses iOS-first vs cross-platform tradeoffs.

11:40 – 15:00

07 · Tradeoffs: Expo vs native SwiftUI

The agent outlines honest caveats of Expo vs native -- chat streaming is the perf trap, virtualisation must be designed in from day one.

15:00 – 17:40

08 · Roadmap generation and final plan file

Agent builds a phased build roadmap and saves it as a markdown file under the plan folder. This is what gets handed to any future agent tab.

17:40 – 18:50

09 · Close and CTA

Wraps up the demo, asks for subscribe toward 100K, tells viewers to comment what they want next.

Atomic Insights

Lines worth screenshotting.

A one-line plan-mode prompt gives the agent one line of context -- everything else is the agent's assumption, not yours.
Plan mode is most useful as a commit step, not an entry point; the conversation before it is where the real planning happens.
Saving the plan as a markdown file means you can hand it to any agent tab, any time, without re-explaining context.
Running multiple agent tabs in parallel is only safe when each tab has the same underlying plan file to read from.
Asking an agent to brainstorm before telling it what to build surfaces tradeoffs you hadn't considered -- which is the point.
The agent will overestimate how long a build takes because it thinks in human-hours; three hours of agent time routinely covers what the model estimates as two weeks.
A good companion app scope question is: what can only a phone do that the desktop version can't -- not, how do I shrink the desktop app.
Expo's native UI components (Expo UI, stable in SDK 52) close most of the gap between React Native and SwiftUI fidelity.
Chat streaming is the one place React Native reliably stumbles on feel -- virtualised message lists must be designed from day one, not bolted on.
The markdown plan file is as much for the human as for the agent -- it's a reference when you return to a feature days later with no memory of where you left off.

Takeaway

Plan mode works best as the last step, not the first.

WHAT TO LEARN

The gap between what you imagined and what the agent builds is almost always a context gap -- and the way to close it is conversation, not a better single prompt.

01Hook and premise

Plan mode is most useful as a commit step, not an entry point; the conversation before it is where actual planning happens.

02How most people use plan mode

A single-line prompt into plan mode leaves every structural decision to the agent, which means the plan reflects the agent's defaults, not your product thinking.

03The method: conversation first

Spending time in back-and-forth conversation before plan mode gives the agent the same mental model you have -- after which plan mode produces a plan you'd actually approve.
Discussing tradeoffs explicitly before asking the agent to plan forces the decision to be made consciously rather than by agent assumption.

04Why save the plan as a markdown file

Saving the generated plan as a markdown file means any future agent tab can pick up exactly where you left off without re-briefing.
When running multiple agent tabs in parallel, the shared plan file is what keeps them from diverging.

06Live demo: planning the Pluto companion app

The feature that makes a mobile companion valuable is what only a phone can do, not what the desktop app already does -- that framing naturally scopes the MVP to push notifications, approvals, and quick capture.

07Tradeoffs: Expo vs native SwiftUI

Agent estimates of build time are calibrated to human engineering hours -- what an agent describes as a two-week build is often achievable in two to three hours of agent time.
Chat streaming in React Native is a known performance trap and must be architected for token-by-token updates from day one, not optimised later.

Glossary

Terms worth knowing.

Plan mode: A feature in Cursor (and similar tools) where the agent generates a structured plan for a task before writing any code. Acts as a dry-run step before execution.
Agentic engineering: A coding workflow where an AI agent handles most of the implementation autonomously, with the human acting as a director or operator rather than a line-by-line author.
Expo / Expo UI: A cross-platform React Native framework that lets a single codebase target both iOS and Android. Expo UI ships native-feeling components that map to UIKit and SwiftUI primitives.
Convex: A reactive backend-as-a-service that keeps client data warm via live subscriptions, making app UI feel instant without explicit cache management.
Human-in-the-loop: A design pattern where the agent pauses and requests explicit approval before taking high-stakes actions such as sending an email, spending money, or deploying code.
Sub-agent: A secondary agent instance spawned by a primary agent to handle a task that would otherwise block the main thread -- analogous to offloading work to a background process.
WorkOS: An authentication and user-management API platform used here as the auth layer for the companion app.

Resources

Things they pointed at.

06:05productDepot CI ↗

02:30toolCursor ↗

07:00productPluto

11:40toolExpo ↗

11:40toolConvex ↗

11:40toolWorkOS ↗

Quotables

Lines you could clip.

03:30

“The model is going to be making a lot of assumptions for you and that's just not good product experience.”

Clean, quotable critique of the default plan-mode habit→ TikTok hook↗ Tweet quote

06:00

“These plan MD files are for me. They are for me to remember what's going on.”

Reframes a common debate -- the file is for the human, not the agent→ IG reel cold open↗ Tweet quote

16:50

“You do the plan one time and then you build and it's nothing like what you want it to be.”

Tight punchline that validates the whole method→ newsletter pull-quote↗ Tweet quote

14:00

“The agent will overestimate how long it's gonna take because they're thinking of human working time.”

Counterintuitive and specific -- clips well standalone→ TikTok hook↗ Tweet quote

The Script

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

00:00Planning is one of the most crucial things that you can do when it comes to building features and apps using AI. Agenetic engineering can only be successful for you when you, the operator, have planned properly. Now when you think of planning, you automatically think of plan mode, and you've probably seen your favorite influencer builder talk about I don't use plan mode.

00:18But here's the thing, the way I do planning which incorporates plan mode has yielded me such success in building features, building apps, whether it's client work, whether it's job work, whether it's my own projects, whether it's Pluto. It's been a success for me. And that's what I'm gonna show you in today's video.

00:35I'm gonna show you how I plan. And after watching this video, you're going to be a great planner, which is gonna help you build great features. Let's get into it.

00:42Now real quick, you're probably like, where is this guy at? I'm gonna show you a couple pictures. I somehow ended up in the Sarhant office.

00:49If you're not familiar with Ryan Sarhant, then you're probably not going to understand. But I'm in the office. I'm in the podcast studio.

00:54I don't know how I got here. I give all glory to God. My life has just been insane the last couple of months.

01:00All glory to Jesus. Let's get into the video. So this is how most people plan.

01:05They use plan mode and they'll prompt saying build x for me, they'll turn plan mode on, it plans, they don't even read it and they just build.

01:13Right? And a classic example of this is I can open cursor, I can open a new tab, and I can say, hey. Build me a new dashboard page that's going to contain analytics for all my usage.

01:25And then they'll toggle plan mode. Right?

01:29Plan mode and then hit enter. And then what's going to happen is the agent has this one line of context. Right?

01:34It has this one instruction. And, obviously, if you're using a good model like Opus four eight or GPT five five, you know, the thinking and all that stuff, it's going to help you generate a decent enough model. But in my opinion, this is not enough.

01:45The model is going to be making a lot of assumptions for you and that's just not good product experience, right? You should have an idea as exactly as to what you wanna build and if you don't have that idea, my methodology will help you a lot.

01:57So how do I go about planning? My way of planning is very different. I actually have multiple back and forths with the agent.

02:06Once I'm confident enough with what the agent wants to build, like it understands what I have in my noggin and it has an idea of what exactly I want to build, I also discuss trade offs, right? What if we do this? What if we do that?

02:18Then after a long discussion, then do we enter plan mode. And once we've entered plan mode, I don't get it to build right away.

02:27I get it to create an m d file and a plan folder that I have then I get it to build. Now, a lot of people might be like, oh, what's the point of these m d files? Some people are saying, these are useless.

02:36These m d files, these plan m d files are for me. They are for me to remember what's going on. Because here's the thing, I usually work in parallel, like, meaning, let's say I have this tab working, I have another tab working.

02:50Let's say I have, like, two tabs working and, like, the agent's cooking, like, I can, you know, start playing Among Us or watch a YouTube video. But let's say I'm just I wanna keep on going. What I'll usually do is in another tab, I'll do the planning.

03:01And, actually, I said I have two tabs open. Usually, I have one tab building a feature and then I have multiple tabs planning. Right?

03:08I have multiple tabs planning. Now why don't I show you a great example. I am gonna go to the files and I am gonna show you my plan folder.

03:16So I have a plan folder here and under plan I have active, apps, desktop, and simplified. So if I click on, let's say, active, why don't I actually, I'm gonna click on apps. This is a feature I'm working on.

03:27I basically want people to be able to set up any sort of CLI on their Pluto instance, like, very quickly, like an app store, you just download it. And this is the plan I generated.

03:38Now this is not your standard plan mode plan. This is a plan that was generated after tons of back and forth. Not only me in the age of a me and one of our engineers at our product studio.

03:49Right? So these are the type of plans I want because a, I know what's going on and what it's going to build, and b, I'm giving the agent the necessary context it needs. Right?

03:58So instead of having these weak plan modes where you just tell it what you want and hit plan mode, you need to start having a conversation. And in that thread sort of built the context for the agents who would understand exactly what it is that you wanna build.

04:13And instead of just yapping, why don't we build out a feature together? But before we do, a quick word from today's sponsor.

04:19I have a quick question for you. Name me something slower than GitHub actions.

04:27You can't. And frankly speaking, you and I deserve better. We deserve something faster, something better, something that agents can use.

04:34And I found that solution and that is the Depot CI. Depot CI is a CI engine that we deserve.

04:41Listen. It's used by companies like Posthog, like Jane Street, like big money Jane Street. Listen.

04:46Not only is it much faster than your lousy GitHub actions, but it's fully programmable. You can get your agent to set you up in just a few commands and you're ready to go. You don't even need to write code to get started, but I can already hear your question.

05:01But, Mike, how much does it cost? How much is the pricing? Look at what it cost.

05:06It cost $0 dot $0.00 $0.01 per second compute.

05:12Now I didn't understand this. So let me quantify it for you simply. So a standard junior chicken, according to today's price, inflation is crazy, is $3 and 69 to $3.89.

05:26That means for the price of one junior chicken, $3, you can get over five hundred minutes. That's over eight hours of compute. Listen, guys.

05:35Don't get stuck with mediocre. Don't get used to mediocre. You deserve something better.

05:39Make sure to check out Depot CI. The link is in the description down below. So what you and I are gonna plan is most of you are familiar with Pluto right now.

05:46It is basically my agentic platform platform for businesses. There's a bunch of stuff you can do. You have a dedicated chat tab.

05:53You can set up tasks. You set up routines. You can give your agent a specific email.

05:58There's thousand plus connectors. You can have files, index, OCR, all that stuff here. Right on the dashboard, you can connect via iMessage, Telegram, Slack.

06:07You can give your agent its own card. I ordered shawarma with this the other day. You can give it a phone number.

06:12You can connect your bank accounts. And not only that, you can also talk to it. Hey, Pluto.

06:18Hey. You can do stuff like that. So there's a lot of cool stuff with Pluto.

06:22Now what I want us to plan together is a companion app. I want a mobile companion app. Now I don't want it to be one on one with the desktop app because it's just impossible or the web app.

06:34So let's build this companion app together. So I'm gonna say, hey. So I want to build a companion mobile app for Pluto, for people who use Pluto.

06:43Now I don't expect it to have all the features and be one to one with the desktop or the web app, but I want it to be useful enough where people can use it, you know, chat with it, get notifications, and all those type of things. I want you to brainstorm some ideas on what a great companion app would look like and share with me your thoughts.

07:02So I do this. I don't do it in plan mode. We are having a conversation.

07:08Now, if the model gets jumpy and tries to build, I'll tell her to shut up. I'll tell her to stop. But right now, what I want is for me and the model to chat and to plan.

07:18In the meantime, y'all hear about this, man? Y'all hear this broke my heart man.

07:23How is JJ gonna leave the sideman? How am I gonna watch sideman among us and there's no JJ doing dumb stuff?

07:30Like, I'm genuinely heartbroken. I thought it was a prank at first but it's not.

07:35If anything, this is actually how I feel. Let me play this video for you all real quick.

07:39It's funny by Harry. Important thing in life will always be the people in this room right here,

07:46right now. Salute me if I'm here.

07:58Foul. Foul. Foul.

08:00Foul. But yeah. I'm a miss JJ being the sideman.

08:03Alright. So our plan, our first iteration has been done. Well, it's still streaming, but it's basically done.

08:10Let's start reviewing from the top. Um, it actually is asking me some questions. Let's review first from the top.

08:15So it's saying the mobile is the agent inbox remote control, not the cockpit. Yep. That's fair.

08:21Don't try to shrink the dashboard. The companion app jobs to reach me. Let me greenlight the agent.

08:25Let me talk to it. Keep me aware. Everything flows.

08:28The killer moments, your differentiators. These are the things that make the app genuinely useful to use daily, and they map directly to data models you already have. Approved from lock screen, the hero feature.

08:38You already have agent card approval requests, amount merchant purpose expires at a push like Pluto wants you to spend $42 on OpenAI a oh, that's fire. Approve and deny.

08:49See, okay. We're talking. Action draft proposals.

08:52Action draft, basically, I explained this in a previous video. I think it's called Claude Cowork. My agent's better than Claude Cowork or whatever.

08:58I'll I'll put it right here. Basically, when you communicate with the Pluto agent, there is approve, there's draft, and there's block.

09:06Approve, meaning it just does the task. Draft, meaning it asks you, like, like, for approval. Like, let's say, it's sending an email or spending some money.

09:14Like, that should be human in the loop. You should check for that stuff. The third one isn't actually just blocked.

09:18The third one is deploying a sub agent. So let's say, it analyzes the task and say, okay. This is a heavy task.

09:23This will block the main agent. So let's deploy sub agent. And the final one is block.

09:28Right? If you're doing, like, some weird or nonsense things, it's not gonna do it for you. So that's great.

09:32It then says task done blocked needs you. Right? That's great for the Kanban task.

09:36Daily briefing routine deliveries. I love that. Voice first capture, quick capture chat on the phone push to talk.

09:43I love this. So it says what I put in MVP version one, auth and device pairing, chat, push notifications, approval inbox activity feed, agent status poll.

09:54Okay. Pill. So that works.

09:55And then fast follows, task routines files, and home screen widgets plus live activity. So one thing that, uh, the models do is they overestimate how long it's gonna take because they're they're thinking of human working time.

10:07So it'll like be like, oh, this will take two weeks to do well. No. We're doing it right now.

10:11Three hours. Two hours max. The native magic layer, share sheet extensions, this to start.

10:16Okay. This looks good to me. Convex is the reactive spine, chat transport, Work OS is gonna work.

10:22Okay. It's a suggested phasing. Push, infra, approvals, chat activity feed, then task routines, control.

10:28Okay. That's fair. Share sheet, Siri app intents.

10:31That's cool. Now it's saying here, OpenDecisions, want you to call iOS first or both? Expo gives us both, but polished budget differs.

10:39Is the agent card approval flow live and central enough to be a hero? He said yes. Multi deployment users, that's fair.

10:47Appetite for building, push in for now. Here's the thing. Now, do I build a companion app that works both on Android and iOS, or do I just fully stick to iOS?

10:58Because here's the thing. With Swift, the Swift UI just looks great.

11:03It feels native with Expo. Polishing is gonna take a lot of time, but at the same time, I have to ask myself, you have to ask yourself, are Android customers worth it? I don't know.

11:13Like, part of me wants to build a native app just, you know, but, like, there are people who use Android. But then at the same time, like, are my type of customers, right, businesses and their employees, do they have Android phones? Yeah.

11:25Yeah. They do. They do.

11:26Right? So I think we might wanna do a full on expo app. So how do you wanna take this forward?

11:32Write a full PRD product spec, write a phased implementation architecture plan, focus on I want to keep chatting. Actually, we're gonna keep chatting. One thing that I want to discuss is if we build an Expo app, can we make it feel and look native, especially on the iOS front?

11:49I really like SwiftUI, and I know there's Expo UI, is basically the native port. One thing I really wanna make sure is this app looks, feels native, is fast, quick clicks.

11:58Can we make this happen? So, again, it wanted to jump from to me, you know, saying go build, go plan, whatever. We're not done yet.

12:06We're gonna discuss. So we're gonna do cross platform and yeah.

12:11Again, we're gonna have this back and forth. Right? But like I showed this diagram here, I give agent response.

12:17I give agent response. I give agent response. Once I'm happy, then we enter plan mode, create the MD file, and then we'll build the thing for building.

12:25If you all watched my last video, you should check it out. It did really, really well. A lot of people enjoyed it.

12:30So make sure to check out that video. Now going back to, uh, cursor, it is cooking, so we're going to let it do its thing.

12:39Alright. We got a response. This is great question.

12:41Why native feel is real now, not hand waving? The thing you pointed out, Expo UI just went stable in Expo SDK. That's great.

12:49A few details that matter to you, Expo UI API was renamed to mirror Apple's toggle. The okay. So we can get the native feel.

12:55You're never locked out of SwiftUI. Native notification is native now. Widgets live activities are stable.

13:01So this is great. Both platforms actually make this better. I love this.

13:04Says how I guarantee fast quick click feel. Tap feels instant because Convex data is already warm via live subscriptions. This is why you should use Convex, native gestures, native press.

13:14So it's looking good. The honest caveats, especially for this app, chat streaming is the one real perf trap. A native implementation re renders the whole message list on every token and will feel janky.

13:24The one place React Native's app reliably stumble. Mitigation is known and not hard.

13:29Batch details that this is virtualization. So but it has to be designed from day one, not bolt on for Pluto. Chat is the core surface.

13:36I prototype first to derisk. Fair. It's not automatic.

13:39A lazy expo app still feels like a web view. Native feels deliberate posture. Okay.

13:43Okay. You're fair. You're fair.

13:44You have to maintain two implementations for the fanciest bits. I know.

13:48I know. The trade off stated painfully. A full native SwiftUI plus full native Compose is still the absolute gold standard for Fidelity, but it's two x to work and fragments you from TypeScript to Fax.

14:01Given that you want both platforms or want this, it's basically telling me to stick to Expo. Okay. That's great.

14:07Is there anything else I want to chat about the app? Again, I want it to be a simple companion app. I want it to be multiplatform.

14:14I want it to feel fast. The one thing I want to figure out is the structure and the road map of which our building.

14:21Okay. So I think you understand all the context of what I really want. I would like for you to build out the road map on how we're going to build this, like step by step what comes first, what comes second, what comes third.

14:33Now, to a lot of you, this might be like, oh, you're spending all this time planning. When do you get to building?

14:39You see, because I spend all this time planning, when I when the AI builds it, it actually builds it close to what I want it to be. You do the plan one time and then you build and it's nothing like you want it to be. So this little sacrifice is very, very important.

14:54And while I'm doing this, I could be planning something else as well. But the focus is on this one so you can understand my thought process. Also, while we're waiting for the egg to cook, if you can this is the sub count right now.

15:05If you could do me a solid and subscribe to the channel. We're so close to a 100 k. I would love that 100 k plaque.

15:11It would mean the world to me. So if you go on, if you haven't subbed, make sure to sub. A lot of you don't sub.

15:16If you sub, you're gonna have a great day. So make sure to sub. Alright.

15:20It's done. So we see this m d file that it created. Let's check it out.

15:25Now, notice how it created under my plan folder in mobile. I don't know if this is something I added in my agents. Md, I did not.

15:35But it just knew based on all the conversations I've had building this app that this is what I want done. So it built out the roadmap, the plan file. So let's see, Pluto Mobile Companion build roadmap.

15:47Goal, to ship native filling iOS, blah blah blah. Here's the diagram, tech stack, guiding principles, and in the phases.

15:56Right? So first, it wants to prove the three riskiest integrations on a real device before building features on top. I respect that, making sure auth works, making sure the streaming works, making sure the pushing works, and that it feels native, and then we build out the rest of the stuff.

16:11Now here's the thing, this is where I would hit plan mode now, but it almost knew that this is what I've been doing. So what I want to do is I'm going to say, alright. You have all the context.

16:20You have all the entire road map. I want you to finally build the ultimate plan to build this to completion. And then we're gonna enter plan mode and hit enter now.

16:31It already did the thing, the step, but I just wanted to show you. I would do this. It would generate the plan, and then I would tell it create the markdown file in the plan folder under a specific folder just so I know.

16:44Because what happens is I might not build this today, but I want this plan here. So when it's time for me to build this feature, I know exactly where I left off, what it is that I wanted to do, what it is that I was thinking.

16:57The MD files are just as much for you as they are for the agent. Oh, look at this. We have sub agents deployed.

17:05We're using Composer two five fast, which is a great model for these type of things. Great model. Very cheap.

17:10By the way, I'm about to put y'all on a new track while, uh, Cursor continues to plan.

17:33Alright. Before I get kicked out, we should stop. Alright.

17:35So the plan was generated. Let's look at this. It says builds a mobile companion app, new ConvexClient, blah blah blah, scope and assumption.

17:43Alright. It's telling us to stack. We have the beautiful diagram.

17:46We need the diagram. Helps us understand, giving us even, what chat streaming could look like, the different phases.

17:54Yeah. And then the to dos. The to dos is what I like for from plan one because this helps the agent.

17:59Right? So now I can have this rebuilt. Right?

18:03Or I usually tell it to create the markdown file, but it already did that because it knows me. Isn't it great when your agent just knows you? So I have this right here under mobile, and this is how I plan.

18:16Now this, you can give this to you can give this to Codex, you can it whatever agent you want. It has all the details exactly how I want the app to look, feel, and all that stuff.

18:26So once the agent has built this, even if it's not perfect to what I imagined, it's pretty darn close. Invest some time and some planning.

18:36Invest some time in talking and going back and forth with the agent. Thank you all for watching this video. Make sure to like, comment, subscribe.

18:42I hope you enjoyed it. I hope you learned something new. Let me know in the comments down below what it is that you want me to record next.

18:47I appreciate you. I love you. Thank you.

18:49Peace.

The Hook

The bait, then the rug-pull.

Most agentic coders treat plan mode like a shortcut -- type one line, hit enter, watch the agent guess the rest. Ras Mic built his entire product studio on a different premise: that the conversation before plan mode is where actual planning happens, and plan mode is just the step that writes it down.

Frameworks