§ 01 · The Hook
The bait, then the rug-pull.
Teresa Torres doesn't use Claude to write for her — she uses it to run her life. Two terminals, one note-taking app, and a 3-layer context system that took months to build incrementally. The result: a 9,000-word blog post in one and a half days, a morning briefing from one typed word, and a system that gets smarter every session instead of starting from scratch.
Who's talking.
00:00hostPeter Yang
00:08guestTeresa Torres
Where the time goes.
00:00 – 03:00
01 · Intro: beyond coding
Peter introduces Teresa as a product coach who uses Claude Code for everything — not just coding. Teresa shares her dual-terminal + Obsidian setup.
03:00 – 07:00
02 · The `today` command
Live demo: Teresa types `today`, Claude hits Trello, scans her tasks folder, rebuilds today.md, runs a research digest. Full morning briefing from one word.
07:00 – 14:00
03 · Task and idea inbox system
How Teresa's tasks folder works: markdown files as tasks, Obsidian front matter, due dates, tags. `new task`, `new idea` commands. Inbox sync from phone via Obsidian Sync.
14:00 – 21:00
04 · Blog post drafting in plan mode
Live demo: rough notes → plan mode outline → Claude proposes structure → Teresa explores alternatives → SEO keyword research mid-flow. Claude as sparring partner, not ghostwriter.
21:00 – 26:00
05 · SEO and alternative structures
Claude does live Google searches for competing articles and keyword volume. Teresa uses this after drafting to tune subheaders, not before.
26:00 – 30:00
06 · Writing with Claude, not delegating to it
Teresa writes every word. Claude reviews section-by-section: what's working, what's unclear, technical accuracy, typos. Having Claude ask 'ready for phase two?' keeps momentum going.
30:00 – 36:00
07 · Context window management
Teresa doesn't let Claude auto-compact. When context is full: Claude writes process-notes.md, she clears manually. She's building a sub-agent documenter to automate this.
36:00 – 44:00
08 · The 3-layer context system
Layer 1: global CLAUDE.md (short, always on, working preferences only). Layer 2: project CLAUDE.md (per-folder rules). Layer 3: reference files (business profile, target audience, differentiators). Claude only loads what's relevant per task.
44:00 – 49:37
09 · 3 tips to get started
1) Whenever you explain context you'll need again, capture it in a file. 2) Work-vs-personal split is the minimum viable structure. 3) Ask Claude at end of every session: what should we add to context?
Lines you could clip.
01:00
"How do I pair with Claude on everything that I do?"
26:30
"I wrote a 9,000-word blog post in one and a half days. There is no way I would have done this myself."
32:50
"When the context window fills up, bad things happen. Claude gets dumber."
39:40
"I don't write my CLAUDE.mds anymore. Whenever we're done working, I say: what did you learn about working with me? What should we add to the CLAUDE.md so this goes smoother next time?"
33:00
"Using Claude Code right now is living on the edge of what's possible today. And if we're gonna build AI into our products, we should be living on that edge."
47:00
"I forced myself all day every day, every time I do a new task, to think: how can Claude help?"
Things they pointed at.
Word for word.
HOOK opening / re-engagementCTA the pitch analogy
00:00HOOKAt the beginning of my day, I literally just write today, and we can see what it does. It generates my to do list for today. It's gonna go to Trello and see if anybody on my team has added any new Trello cards. And then it basically runs this Python script, and then you can see it's gonna read my today m d, and it's gonna update it, and you can see it popped up here. So this is a link to my research digest, my blog post. It's 9,000 words, and I wrote it in one and a half days. There is no way I would have done this myself. That is insane to me. Here's how I'm thinking about it. What do you think? Oh, I don't really like that. What if we try this other way? And then I do that throughout the whole construction of the blog post. You can just be like, interview me about my business, and then it will ask you a bunch of questions, and then it will write a context file for you. This file we're looking at, this profile, I never wrote a word of it. Every time I have Claude add a context file, I say what index needs to be updated, and it just figures it out. Let's say I'm just overwhelmed by all of this stuff. What are your three tips to get started? Whenever you find yourself explaining context to Claude, stop and think about, am I ever gonna have to explain this context to Claude again?
01:01Okay. Welcome, everyone. Uh, my guest today is Theresa Torres. Uh, Theresa is a legend in the PM field. She's author of Continuous Discovery Habits. But today, we're gonna talk about how Theresa uses Cloud Code for everything from writing to editing to task management to coding. Really excited to get her to give us a tour of her Cloud Code projects and show us how it's actually done. So welcome, Theresa. Thanks for having me. I'm excited to do this. Yeah. Um, so, you know, I've also used Cloud Code quite quite a bit, but I I'm super interested in seeing your, uh, high level workflows. Maybe you can give us a quick tour first. Yeah. Let me share my screen.
01:38Okay. So what we're looking at is really simple. I have Obsidian here on the left. And then and if people aren't familiar with Obsidian, it's just a note taking tool that's based on markdown. And then I have two terminal windows on the right. And the big thing that I use ClaudeCode for, like, I started using it coding just like everybody else, most that's most people's entry into it.
01:58But I what I really loved about CLOD Code with coding is it's almost like your pair programming with CLOD. And I was like, wow. How do I pair with CLOD on everything that I do? And so I started doing this for writing and for strategy and for literally all of the work that I do. This is my setup. This is my work Obsidian vault. So I have an Obsidian vault called work.
02:21In it, there's a bunch of subdirectories. One of them is LLM context, which I'll get into that. And then my notes, which is just any notes that I have. I have a research project that we can talk through. I have my task system. Worthy reads are like articles that I've saved that other people wrote that I really like and then my writing directory. And then you can see I have a clot m d file here, and we can kinda go through all of this. But I'm gonna start with tasks. So the way my task system works is I have folders for different things. So I have a folder for bugs. Have I a folder for ideas. I have a folder for
02:54tasks. And every day, I launch Claude inside my tasks folder. So you can see this top terminal window. I'm in tasks. I'm just gonna launch Claude. And at the beginning of my day, I literally just write today and we can see what it does.
03:09The first thing it's gonna do is it's gonna go to Trello and see if anybody on my team has added any new Trello cards to my board. You can see good news. There's no new Trello cards. I've already run this today, but we'll just see what it looks like. And then it basically runs this Python script where it generates my to do list for today. So that's what we're looking at here in Obsidian. It says Monday, November 10. I have one overdue item. I have a number of items that are due today, and then I have my in progress ideas that I'm working on. And
03:40this script just generated this file. Every task is a markdown file. So you can see in my tasks folder, I just have a whole bunch of tasks. Basically, my this task that's overdue was due on Friday. You can see it's just I have to call the grooming place for my dog and schedule an appointment for her, and I didn't do that yet. And then all my other tasks, like, have calls with a couple people today.
04:04I'm writing a Claude code safety article. I've gotta add some images to it. All this stuff just shows up on my to do list every day based on what I put in my tasks. Got it. So the other tasks, like the calls and stuff, is coming from a calendar or is it coming from just your notes? It's tasks that I've created. And I'll show you. We're gonna create a new task together. I'll show you what that looks like. Got it. So that's the first thing is when I run the today command, I do it every morning. It's basically looking through my tasks folder
04:32and looking for anything that's due today or anything that's overdue. And then it creates this today file for me, and it's just telling me what do I need to do today. And then I distinguish between tasks that have a due date and ideas that you know, they're, like, more like projects that don't have a specific due date, but they're things that I have going that are ongoing.
04:53So you can see, like, I have a podcast just not possible. That's, like, always in my in progress because it's reminding me to, like, make sure I schedule guests. And you can see here I have a single episode. This is my project for, like, getting that episode ready to be released.
05:11And then I have a task for social media. But this is really just my, like, remember to schedule guests kind of thing that's always in progress. And so what this does is it gives me a lot of flexibility. If I just have a random idea, I can just say to Claude new idea and it will tell it what it is. It will create a markdown file. It'll put it in my ideas folder.
05:34If it's a task, I can give it a due date and then it will show up on my on this like to do list. This is just called my today MD on the day that it's due. Should we try making maybe like record pockets of Peter or something like some sort of task? Yeah. So we'll do a new task in just a second here. The other thing that's happening when I run this today command,
05:56it is I've built this research system. So I'm trying to keep up to date on like academic research that's related to my work. And so I built this system that searches a preprint server and then Google Scholar every day. And it gives me a research report of what's relevant to my work, and that gets added to this today list. So you can see right here, it's not here yet, and it's because we haven't done that part. So it's running this slash command that's part of my research project,
06:25And now it's asking me, oh, it wants to go see if there's a research queue, so we're gonna say yes to that. Mhmm. And then you'll see it's gonna update this today list with any research for me to review. The queue is empty. It's creating my research today MD.
06:42Got it. So this is just like searching some websites to find new research. Right? Yep. And then you can see it's gonna read my today MD, and it's gonna update it. And you can see it popped up here. So this is a link to my research digest. So That's great. That that's my today command, and you can see it gives me a summary. Like, this is what I just did. I created your today m d with the overdue task, 10 in tasks for in progress, your research,
07:06whatever. And then you can see it's also creating a this week view and a next week view. Got it. Okay. Yep. Okay. So let's say I wanna do a new task. We're actually gonna write a blog post to that together. So I'm gonna do new task, Write plan auto accept Claude
07:25blog post. Do today draft outline
07:32with Claude. And it basically this this tasks folder is set up with a Claude MD so that Claude knows what to do with this. It knows to create a task. You can see it's using Obsidian front matter, and I've defined this whole system for Claude so it knows to do this. It's defining it as a task. It's putting the due date as today. It's tagging it for me. It's gif basically giving me kind of a checklist. I it's actually not supposed to do this because it's kind of a crummy checklist.
07:58It's supposed to just take my notes, but pretty good. We're gonna go with that. And then let's see if sometimes, Claude is smart enough to put something due today on my today MD. It's not this time. So I'm gonna say add to today MD. Got it. Okay. And then what that's gonna do is it's just gonna add that task. You can see it's here. So now I have my right plan auto accept mode. What I love about this is I work out of the terminal most of the day and I keep this tasks terminal open all the time. So if, like, I think of something that I have to do or I have a random idea, I literally can just be like, new idea, blah blah blah, and it's done. And then I can just go back to my work in my other window. And it's it's a lot faster than, like, opening a web browser, going to Trello, creating a new card, setting a due date. Right? It's just it's super fast, and that's what I love about it. Got it. This episode is brought to you by Optimizely.
08:50The problem in marketing usually isn't a lack of ideas. It's a lack of time. If your to do list keeps getting away from you, then take a look at Optimizely Oppo, an AI agent platform built specifically for marketing. With Oppo, you can use AI agents for SEO and general recommendations, AB testing, website analysis, and much more. Oppo knows your brand inside and out and plugs into your existing tools and data systems that you can save time on labor intensive feature testing, reporting, and more. See what it can take off your plate at optimisely.com/ai.
09:20Now back to our episode. The other thing I'll share is it allows me to do things like, what are all my marketing ideas? And it's just gonna go search my ideas folder for anything related to marketing. And so, like, whenever I'm working on a new project, I can just ask Clyde, like, what are my tasks related to this? Or what are my what are my idea open ideas that are related to this? And it just goes out and finds them. And, like, Trello has search, but it's not the best, frankly. Okay.
09:50So I like this a lot too. So you have, like, ideas and memories and, like, random re research snippets that you pull. Yeah. So you can see type. Like, in my tax folder, I have bugs that have been filed. I have ideas. Import is for if I'm importing a bunch of stuff from Trello. I wanna review it to make sure it imported correctly. My inbox is how I get stuff from my phone to my to this whole system.
10:13Memories is like things it's like little snippets that I wanna remember so we can kinda look at what's in here. I think this is safe. Like, I have a affiliate link to the AI evals course that I always forget what it is, so it's literally just here's my affiliate link. This is a task. I don't actually know why that's in memories. Real real real quick. What do you mean by inbox? Like, send an email to yourself over the phone, Or or how does it Yeah. Uh, so this this is a little bit of a hack, and I'm actually moving away from it. So I mentioned I moved a bunch of stuff and it broke some things. But, um, I was using Dropbox to sync my vaults,
10:47and I'm now moving to Obsidian Sync to sync my vaults. So the Obsidian Sync, I can use the iOS Obsidian app, and I can just access all of this stuff from my phone. But when I was using Dropbox, I had to set up. This is like a iCloud linked folder so that I could add to it from my phone, and then it would automatically show up here. And then I could pull it into my task system.
11:09That was kind of a pain in the butt. So I moved from Dropbox to Obsidian Sync to make that easier. Got it. Okay. So, basically, Obsidian and Cloud Code runs your life right now. Yeah. That would make sense. Yeah. Pretty much Cloud Code and Obsidian run my life. Alright. So let's go back to the writing the blog post task. So, like Yeah. How's it gonna help us write this blog post? Yeah. So
11:28I could do this. Let's do this. We're gonna get rid of this dumb checklist that Claude made because I don't like it. And we're just gonna do, like okay. I'm gonna talk through I'm just gonna, like, talk out loud while I plan out my blog post. So for context, I'm in the middle of writing this huge Cloud Code series for people that are nontechnical. And it started with, like, what is Cloud Code? Why is it different from Cloud in the browser?
11:52And then the second post in the series was about, like, how to give Cloud a memory, which if you wanna get into that, all my context files are also here in Obsidian. And then the third blog post that's coming out this Wednesday is about safety, so how to use Claude codes safely. And then this one we're gonna start together is the fourth article. And this is where I'm gonna get into
12:11doing projects with Claude. And so I wanna write a blog post where I talk about, like, plan mode, auto accept mode. And then I always wanna use examples in my blog posts that are related to product managers. And so, like, in this blog post, I might think, like, let's talk through
12:30plan mode and auto accept mode if you were planning a new feature. So I'm starting to think about, like, what's my example gonna be in the blog post? And I have an AI product called the interview coach. This is a product where in my class where I teach how to conduct effective customer interviews, we do practice interviews with each other in class, and then the students can submit their transcripts
12:51CTAand get, like, detailed feedback on how good of a job they did. And the thing with this product is it was built for my story based interview coach or story based interview class. And so the coach gives feedback on story based interviews. And my new feature idea is, can I extend it to work with non story based interviews? So this is like a feature idea I have. And so in the blog post, there's a little meta. I'm gonna use this feature idea to show how I would use plan mode and auto accept mode. Right? So Okay. Yep. I'm gonna this is all I have. This is literally the state of this blog post. This is real. I have not written a single thing.
13:31CTAAnd now I'm gonna go over to this terminal window on the bottom and I'm gonna launch launch Claude and this is in the context of my writing vault. And so my writing vault has a Claude MD that talks about how I like to write and how I like to write with Claude. So it has different context than this tasks window. So like in my writing window, I can't type new task. It doesn't know anything about my task system. I have to do that in my tasks window.
13:55CTAGot it. So we're gonna do this. I don't know why it thinks it's a new folder, but we're gonna ignore that for a minute. And then I'm gonna say, I want help creating an outline for a new blog post. I do this in plan mode, which is actually what the blog post is about. You can find my early thoughts in oh, see, I would not this is a little weird. I would not do this in the task card.
14:20What I would do is I would come to writing. I would go to my Claude code series and I would do a new file. Got it. Let's do plan mode.
14:33And that did not go where I wanted it, so we're gonna move it in plan mode. And I actually did not tell Claude enough information because it's not. It's gonna look for plan mode in the current directory, but the plan mode is actually in Claude code. So we'll see what it comes back with. We'll see if it finds it. Okay. So, basically, you have a Claude okay. So just real quick. You have a Claude dot m d for your writing vault
14:56that has a bunch of writing styles and, like, you know, that kind stuff. Right? Yep. Is that got it. Yeah. So if you look at here on the left, like, tasks is a then you can think about tasks as, like, a Claude project. I launched this instance of Claude in the context of tasks, and then I have a writing folder. And I launched this instance of Claude in the context of the writing folder. So the writing folder has different rules than the tasks folder. They each have their own CloudMDs.
15:25So the CloudMD for tasks explains how the task system works. It explains how tagging works. It explains how front matter in Obsidian works. So when I just type new task, it knows exactly what to do with that. Whereas my writing folder doesn't have any of that context. It just knows it. Like you can see here, I told it I wanna write a blog post. It found plan mode. It read it. Now it's asking me to read my style guide. And that's because my CloudMD
15:49in my writing folder says before we do anything together, start by reading my writing style guide. Got it. Okay. So I'm gonna go ahead and say yes, it can do that. Do you ever just ask it to just like skip all the permissions? Because it's kind of annoying to Yeah. You know what? I have so I just moved everything. Everything used to be in Dropbox, and I had all the permissions set up where my tasks could read and edit everything inside tasks and writing could read everything inside writing.
16:15But because I just moved everything, those permissions aren't I literally moved everything yesterday, and so I haven't copied over those permissions yet. I'm learning, like, when you move things in Claude, it kinda breaks things a little bit. So it's asking me which features do I wanna focus on. I actually wanna do both.
16:33What's your intended audience for this post? Product people new to Claude should already know that. That's a little disappointing. Wait. So so, like, is asking you questions because you prompted to ask questions? It asked me questions because it read this plan mode document, and it's now, like, trying to get more information from me to be helpful. Normally, I have a little bit more here. Right? This isn't really enough for Claude to work with yet, but let's let's see what Claude does.
17:00Yeah. See, it's trying to do too much right out of the gate, but let's see what it does. Now I understand the direction. Let me create an outline.
17:12Fine. It's already got a headline.
17:17Introduce plan mode and auto accept mode, what it is, when to use it, what you plan, what it is, the dangers.
17:27That's that's Like great. Yeah. Is kinda nuts. Right? Like, kinda went crazy. Okay. So I'm gonna tell it it went crazy.
17:36Uh, it's kinda like, uh, why I use Cloud Code to, like, write out a prospect. I always wanted to put put the spec in MD files so can delete a bunch of stuff because it it tends to do too much. You know? Yeah. So usually to be honest, usually, I have a little bit more. So let's let's actually play with this draft a little bit. Yeah. I would probably give Claude something like
17:56goals help people new to Claude code do bigger projects.
18:04Claude code. Use planning a new feature as a product manager
18:11as an example throughout. And then we could do I can use my personal experience with
18:20expanding the interview coach. I'm gonna ignore typos for now because Cloud won't care. Mhmm. And then I'm gonna do something like
18:29less about the features, these features, more about how to get Claude
18:40to do good work. And let's see if this helps.
18:46Got it. Let's see what it comes back with. So okay. So after this, I'll, like, gonna ask you to do online research or what's the what can I can do next? Yeah. So you can see already it came back and it's like, oh, okay. I made too many assumptions. Yeah. And it's like, okay. Now, like, now is this, like, the right direction?
19:06So now I might now I what I really wanna do is ask Claude, like, you asked about research. We gotta come out of plan mode for this. Has anyone else written a blog post about this? And I'll go and I'll go search and see what other blog posts are out there. So I'll often do this. I'll often like, before I sit down to write, I wanna know what else has been out there. If somebody else has written this article for my audience, I'm not gonna write it. I'm just gonna, like, write a different article and just link to that article.
19:34I wanna make sure that I'm writing stuff that's unique. And then now sometimes someone may have written this blog post, like, in the context of coding, and that's fine. Most of my audience isn't gonna read that. And so but I can read that article and learn from their experience and then translate it to product management. So I always use Claude code to, like, do searches for me. The other thing I use it for while I'm writing is, like, I'll make a claim and I'll be like, oh, is that claim true?
20:00And then I'll ask Claude, like, can you do some academic research for me and see if there's evidence behind this? Like, is this really a true claim, is it just something that I randomly believe? So let's see what Claude came back with. So a lot of people have written about this. Of course, Anthropic has written about it. They have great documentation.
20:15Most are written for developers, engineers. Yeah. Yeah. Perfect. I think people wrote, like, technical stuff about Plano, but not about this use case. Yeah. So then I might ask, like, okay. Um, how do people search for this kind of content? What keywords
20:32from an SEO standpoint should I keep in mind? I almost never do this to start. I do this at the end, but we can do it to see how it does. Oh, that's very smart. I I I never asked this question from my blog post. So that that's that's that's that's important. Yeah. And, again, it just goes off and does its thing. And I let it do its thing. In the meanwhile, I'll be like over here noodling. I'm like, how am gonna structure
20:54this blog post? And like, this is really stubbed out. There's not really a blog post here. Right? So Mhmm. I might start to think about an outline and be like, okay, well, the introduction needs to be strong hook
21:09around value of doing project work with Claude code. Then introduce, like, Claude modes,
21:20plan, auto accept, reference safety article,
21:28walk through detailed example interview coach. So let's see what it came back with.
21:36So I found it's really good at keyword research. Like, good at keyword research. I tend to do this after I've written the article because I really wanna write my articles for humans. But then I might tweak some words in my subheaders or my titles to to, like, target higher volume keywords. What what what is it actually doing? Is this using Google searches here, or is it actually doing some fancy stuff? Yeah. You can see it's just doing searches and looking for what's ranking well. Oh, is that what it's doing? Yeah. Oh, interesting. Okay. Yep.
22:05Okay. So it's sort of my SEO researcher. Interesting. And then I could be like, I've started to stub out an outline
22:13in plan. Can you suggest alternative structures?
22:20There's not enough meat here. It's probably not gonna do a very good job here, but we'll see what it does. And so I do this. This is literally how I write. I'm, like, doing all my thinking really rough. I'm going back and forth with Claude. I'll have questions like, what do I do? Like, how does this work in Claude? Like, I think I know how auto accept mode works. Right? But then I'll be writing a blog post about it. I'll be like, is that really right? And I literally will just pop over here and be like, Claude, is that are you allowed to do this in auto accept mode?
22:49And so that's really nice too. And I know this is a little meta. I'm writing a blog post about Claude code, But Claude can actually answer questions about anything. Right? So I could be writing a blog post about how product teams do customer interview analysis, and I could still ask Claude. In fact, I wrote a blog post about how about customer interview analysis with AI. And then when I was writing that article, I was constantly asking Claude
23:12to look at research for me, like academic research. Like, what do we know about what is lost when we rely on AI synthesis? What do we know? And a lot of that came from just asking Claude to do the research for me. Got it. This is very much two screen kind of process. Yeah. They're definitely using it as a thought partner and, like, a researcher Yeah. In a way. And so you can see here now Claude came back with, like, here's some other ways to structure it. So we could do example first structure,
23:39problem solution structure, journey structure, like your personal story. And you can see for each it's it's kinda like giving me a rough outline. So what I like about this is, like, when I write, I usually have a pretty fixed idea of what the structure is in my head.
23:55And Claude helps me explore alternatives. So it's like I'm comparing and contrasting in the context of writing, which before working with Claude, I would have literally never done that. And then the other thing that it really helps me with, and maybe I'll pull up my web browser to show this.
24:11Let me think about actually, I think I can show it right here. The article that I'm publishing this coming Wednesday is on safety. So it's how do I teach beginners, Cloud Code beginners, how to run an LLM safely on their
24:25computer. And you can see I'll show you. Ever it's organized by tier. So, like, what's the risk of letting it read files? And then at the end of every section, I include what Clot is doing and how you know what it's doing. So, like, there's these tables of, like, this just indicates if it's reading. There's, like because, like, I'm assuming most of my readers don't know Unix commands. Right? So Yeah. Have a nice terminal. Yeah. And there's no way I would have written this article. I'll show you at the bottom. Look at all these look at what we cover. Like, it's insane.
24:55This blog post is you can see it's, like, 8,800 words. And I created this, like, quick reference commands table. This is like how Claude reads. This is how it searches beyond your current directory. This is how it searches the web. This is how it writes files. This is how it executes code. And like, this is an awesome reference now for people that have never used Claude code. There is no way I would have done this myself.
25:19Claude Okay. So you yeah. I started to generate a list like you can see in my outline. Like, I started to generate a list of, like, commands as I used Claude. Like, oh, this stuff should go in my document.
25:34But as I worked with Claude, it taught me, like, oh, I also use these other commands. And it so this blog post is way more thorough than it would have been if I had just done it on my own. Coauthoring me with Claude along along the way, like, section by section. Yeah. Yeah. And then the other thing that I do is, like, once I get to a complete outline,
25:54I actually start writing. And I'd still do all of my own writing. I want it to be in my voice. I personally like to write, and I have I feel like I have a very specific cadence to my writing that, like, Claude just doesn't really get. It gets close, but not really. And so what I'll do is once I have a really detailed outline, I will write a section. And then as soon as I'm done with the section, I just pop over here and I'll say, like, Claude, I wrote the intro. Give me feedback.
26:19And then Claude can see the same file I can see. Right? So I don't have to cut and paste anything. And then Claude will give me feedback, and then it always finds typos. And it'll just be like, do want me to fix the typos? And I can just type yes. Why don't we try it? Can you get it to fix the typos right here? Like, does it Yeah. Edit in your file? Yeah. Can you
26:36fix the typos in my plan mode? Got it. Yeah. And it's just gonna go and do it.
26:44So you can see here it's showing me a diff of what it's fixing. I don't want it just fixing random typos in my blog posts as I write. So one thing I've taught Claude to do is when we review a section, first it tells me what's working well, then it tells me what could be better or what could be clearer. It does a technical review. Am I saying anything that's technically wrong? And then the last thing it does is it lists the typos that it found and it asks me if I want to fix those typos.
27:10And that's just sort of a safety check. Sometimes like I'm using a word that it thinks is a typo but it's not. And so it allows me to quickly look down like this is kind of an example. This is one that told me after the fact. But when we're doing our section by section review, it'll be like, here's what I wanna change. And then I can just say, change it all. Or I can be like, yes, change it all, but don't change this one. Got it. Got Got So I kinda still keep, like, safety guardrails on Claude before it just starts editing my blog post. Interesting. It's interesting that you still okay. So you still manually write the actual blog post from a detailed outline.
27:41You still write it? You're Yeah. I've experimented with letting Claude write stuff. I don't like the voice, and Claude knows a lot about my writing. Like, I have my entire product talk archives in here. I have taken the time to develop a really good writing style guide. For me, good writing has a cadence that, like, you can hear.
28:01And that's the part of my writing. Like, I just can't get Claude to get a 100% right. And then I also like to write to figure out what I think. And so, like, I don't wanna outsource all of that to Claude because I wanna force myself to do the thinking myself. Got it. Got it. I'm using Claude almost like a sparring partner. So, like, here's how I'm thinking about it. What do you think? Oh, I don't really like that. What if we try this other way?
28:24And then I do that throughout the whole construction of the blog post. Yeah. It's like having a editor and a researcher and, like, you know, just to come there sitting there with you. Right? But, like, full time. Right? It's not like I did everything and then you edit it. We're doing it very collaboratively. And then the other piece that I really like is that the way that I used to write, I used to write like a two to 3,000 word blog post in, like, three or four days. And a reason why it would take so long is I would write a section, and writing is exhausting. Right? And so at the end of writing a section, I'd like, I'm gonna go check my email. And then I'd go get distracted for an hour, then I come back and write another section. But when I'm writing with Claude, as soon as we finish the review of a section, Claude's always like, are you ready for phase two?
29:06And I'm like, yeah. Okay. I guess I am ready for phase two. So just having Claude, like, poke me, I maintain momentum. And so I now a lot of my blog posts are a lot more in-depth. They're a lot more detailed. They're getting longer. Like, you can see this blog post is pretty darn long, but you'll see there's there is literally no fluff. Like, I challenge people. This article is coming out on November 12.
29:30Go read it. You can tell me if you think there's any fluff in it. Like, to me, this is just a very detailed guide for how to work with Claude safely. And I wrote all of this myself. It's 9,000 words, and I wrote it in one and a half days. Wow. Like, that is insane to me. Okay. Let let me ask you, like, a few questions about this process. Yeah. So so you have one window open for writing. Do you just do you ever clear the context or you just keep keep going? And and then you let it That's a really good for you. That's a really good question. When I'm working on a outline, like, when we're developing so, like, in this writing exercise,
30:07we're developing an outline. I try to keep all of it in the context of the same conversation. Okay. There will be times like I'll keep an eye on whether or not it's gonna wanna compact the conversation. I always try to keep Claude from compacting. Like if it looks like that's getting close, what I'll do is I'll be like, Claude, we're gonna run out of context window. Let's write a summary of where we are that you can read in when I clear the context window. And the reason why I do that is I want oversight on how Claude is compacting the conversation. Whereas if you use the compact like, if you just let Claude use its own compact tool when you run out, I find that it loses a lot of context
30:43and detail that I don't want it to lose. So one thing that I've done, I don't always do this with writing, but I always like, when I'm working on a project with Claude, in fact, that needs to go in this blog post. So I'm glad we're talking through this. I have Claude create this document called process notes. And as we're working, if it looks like we're getting towards the end of the compact the context window, I'll stop and have Claude
31:08update process notes. And process notes is just a text file that's like, here's what we did in every session so that I have a history of what we've done and the decisions that we've made. And then sometimes Claude, like, just loses stuff in the context window. I can be always be like, can you just search process notes? Like, I feel like we already made this decision.
31:28And so this is a habit I've gotten into whenever I work with Claude and anything meaty. We're constantly co creating process notes together. And I'm getting to the point where, like, I have a a process for my process notes where I'm gonna write a sub agent that's my documenter that Claude will just call to write those process notes when we're running out of context window.
31:49Okay. Got it. You you've gone to, like, uh, this is, like, inception. You have, like Yeah. Processes for I mean, this is the hard part. Right? Like, when the context window fills up, bad things happen. Like, I think you waste a lot of work. Claude loses details. Yeah. It gets dumber. Yeah. It just gets dumber. And so I feel like it's our job, at least for now, to, like,
32:11this process so that when Claude resets and his memory gets wiped, it can pick up where you left off and you'll be fine. Anthropic has tried to build that in with that compact conversation feature, but it's Yeah. It's not very good. So I like to I like to manage it myself a little bit more.
32:28Well, I think that's the thing with this the the this like, that that's the thing with Cloud Code. It just like, it's it's really kind of made for power users, to be honest, like, who really wanna personalize every everything. Like, it's not super it's not super intuitive to use if you're a complete beginner, but you have to, like, go into the rabbit hole and then it gets more and more powerful. You know what I like about it though for product people is that,
32:50like, using Cloud Code right now and especially pushing the boundaries with Cloud Code is living on the edge of what's possible today. And I feel like if we're gonna build with AI, like as product people, if we're gonna build AI into our products, like, we should be living on that edge. And so, like, even just this stuff we talked about with context windows,
33:09I wouldn't have learned any of that if I hadn't been in Cloud Code all day every day. And so now when I go work on, like, my AI product, when I'm working on my interview coach, I have, like, a depth of experience of understanding how to manage a context window, but there's no way I would have developed otherwise. And so, like, yeah, like, in the long run, is this how we're gonna work with LLMs? Probably not. Like, they'll be tooling around it and the labs will get better at managing the context for us. But, like, as product people, we can't wait for that. We're building AI into our products today.
33:38So I feel like we have to be living on that edge so we can build our products on the edge as well. Yeah. You gotta build on the edge. You gotta learn the lay latest info. Yeah. So let that's kind of I I just wanna kinda talk about the context window thing a little more. So, you know, so so context is, like, everything for how effective these LMs work. Right? So I just wanna talk about the three the three layers of context that you talked about in your previous blog post. Yeah. And you don't have to maybe share your cloud.md file, but maybe, like, talk about maybe start with the highest layer and talk about what you put in there, and then and then kinda, you know. Yeah. So I actually have that article over right here.
34:14Yeah. Yeah. Let's see what state it's in. So this article is life. You can find it at producttalk.org. And in this article, I just talk about, like, the opening hook is this story about how I just got sick of trying to get Claude and ChatGPT to, like, be reliably good. And I realized that, like, one of the challenges is that in order for Claude to be good, it needs to know all about me and my business. It needs to know that I'm a product discovery coach. It needs to know that I write at product talk. It needs to know that my audience is cross functional product teams.
34:43It needs to know that I have a course business and what all those products are. It even needs to know things like who I work with. Right? In order for me to say, like, not only in my task window, I can't I can add tasks to my to do list, but I can also say create a Trello card on Wilena's board. And Wilena is my admin. Right? And so I can be like create a tree or I can even better,
35:06I can create a task file in my system in markdown and then say create a Trello card on Wilena's board using this task. So it's like in my system, I'm creating a task for Wilena and then Claude goes and pushes it to her Trello board. Okay. That only works if Claude has the context like who's Wilena and what's her Trello board and it like what's a task, right? Yeah. So in order to get this like, wrote some pretty lazy stuff, new task, blah blah blah. Like in order for that to work,
35:35Claude has to have this view that we're looking at. It doesn't have images. But I just explained this idea that every conversation starts from scratch. We have to create memory. Memory has three layers. You asked about the three layers. The first one is my global preferences. Mhmm. Let me actually pop over to the web because I can show some of my CloudMD files pretty safely. Okay. So this is my global CloudMD,
35:58the start of it. Yeah. I can zoom in a little bit. Perfect. Yeah. Um, this top section I write about in the blog post is just a little thing I learned is that like I don't write my Claude MDs anymore. Whenever I'm working with Claude,
36:13when we're done working, I say, hey, Claude, what did you learn about working with me? Like what should we add to the Claude MD so this goes smoother next time? And that's been awesome because I don't have to maintain this file anymore. Claude does. But you can see, like, I've just defined some personal preferences. Like, I always wanna plan before Claude does anything. I never let Claude just do stuff.
36:34And I do plan at multiple levels. You could saw that with the blog post. Like, I start with just really rough thoughts. We work our way to an outline. Once I have an outline, we start writing. So there's just this this file is across all my projects, coding, writing, tasks, everything.
36:50I just want Claude to know, like, this is what I like. And it's not very long. This file doesn't go much further behind beyond this. It's pretty short. I just have a little section on what how I prefer to get feedback from Claude, and it's really simple. And this lives you can see here, it lives in my local user directory. So on my Mac, it's like tilde slash t torres and then dot ClaudClaudMD.
37:14Yeah. Do you put, like, Theresa as a product coach and, like, you know, here's my business and all that kind stuff here? I'll I don't put any of that here because this file is gonna get loaded in the context window every single time you use Claude. Got it. Yeah. You only wanna put stuff in here that you always want it to follow. And so like if I'm using Claude to like brainstorm Christmas gifts for my husband, Claude does not need to know about product talk.
37:38So I don't want that here. Right? No matter what this short. Yeah. Yeah. This is literally global. So you wanna keep it short. It's gonna go in every single context window. And like I asked Claude all the time, can my dog eat this food? Like Claude doesn't need to know about product talk for those stupid queries. Right? So Okay. Got it. Bit like, what's really important is you have to keep the context window as clean as possible.
38:02Yep. Right? And so this literally is just my global rules. But you'll see later, we just scroll down, I have project specific instructions. So this is actually my writing ClotMD.
38:13And in my writing ClotMD, I have like it says at the start of each section, read my writing style guide. This is actually old. I moved things out of Dropbox. But if you remember, if I go back to the terminal, which is hard to do with this sharing thing here,
38:30when I ran sorry. Here, when I said let's work on a blog post together, the very first thing Claude did was it went and read my writing style guide. That's right. Yep. And it's somewhere in here. Yeah. And that's because here it says at the start of every session, read the writing style guide. Okay. And then this is where I get into my rules about writing.
38:52So Claude is a thought partner, not a writer. You're acting as an editor. You can do some research and development. And this document is also not very long. It's really just telling Claude this is how we write together. Okay. One thing that is in my global Claude MD,
39:08which I don't let's see if I have a screenshot of it somewhere, is this idea of context. So I have my I have my context files defined.
39:17So I've created little markdown files, and the key is they're little. So I have a lot of them because I want to I wanna mix and match them
39:28so that I can tell Claude which files matter for the task at hand and it doesn't have to read all of my context every time. So I have a business profile that just is like, this is what my business does. Here's where you can find my product descriptions. Here's where you can find my marketing channels. Here's where you can learn about my team. I can show you that one. But it basically is telling Claude where to find more details in all these files. And so I think this blog post has a screenshot of that. Yeah. So this is my global CloudMD. We already looked at this top part.
39:59You can see at the bottom, I'm telling it, here's where you can find reference context files. And I basically tell it specifically, only use these files if they're relevant to what I'm asking you about. So if I'm asking you about something related to my business, go look at my business profile.
40:16And if I'm asking you something personal, go look at my personal profile. And then you can This makes sense. So you don't have to, like, load a lot of shit into the default context. Yeah. So you don't crowd crowd the context. Exactly. So, like, if I ask it, I have a product file all about my story based customer interview course. So if I ask it like, hey. Let's work on the landing page for my story based customer interview course. It already knows where to find details about this course. It already knows details about my company. It already knows details about my target audience. I don't have to tell any of that stuff.
40:45And it's because I've given it an index of context files so it can just go pull in the relevant stuff. Wow. Okay. But don't you feel like there's gonna be, like, too many files to manage? Or, like, I guess you get you just you just make Cloud up update your context files. Right? Is is that Yeah. You do? So I do. In this article, I talk about this by the way, this article has a really fun use case of, like, how to use Claude to do a competitive analysis.
41:09So you can see here I'm having Claude, this isn't for me. I used eleven labs as my as if that was my company. And it's generating like
41:18let's see if I can find it. I don't know if I'm gonna find it. But here's the goal. Like, the key is to give Claude just enough to go find what it needs when it needs it. And part of this article is about like, okay, well, how do I maintain all this stuff? Like, I have a ton of files now. How do I keep it current? And so what I do is I have this section about like, okay, how do I keep it up to date? The way we're gonna keep it up to date is at the end of every session with Claude, I literally say, what did you learn about me that we should add to a context file? And then we have a conversation about where to add it. Should it go in a Claude MD? Should it go in one of these LLM context files?
41:57Like, what's the right place? And then I'm working with Claude to make sure like, Claude wants to jam everything into your CloudMD. You don't want that. That's getting loaded in every conversation. So I work with Claude to be like, hey. It seems like that should go in my marketing channels file. Why don't we add that there? Got it. So I think about it as, like, information about me and my business goes in a context file, whereas my working preferences,
42:21like, this is how me and Claude work together, goes in a Claude MD. So, basically, yeah, this is a really important tip. Like, after every conversation, you ask Claude how can you make the context better, and you have all your context in the LM context folder with a bunch of sub subfolders and files. Yeah. Right? Yeah. Wait. So how much are you paying this
42:41cloud employees? They sound pretty useful. How how much are you paying? Yeah. $20 a month? Yeah. I'm on the $100 a month plan. Okay. But that's So you can see I'll show you here in in Obsidian. You can see all my context files.
42:56So, like, for my business, I have a target audience file, a marketing profile, a list of my differentiators, my company overview, my business model. I didn't, like, sit down and create this all one day. That would have been tedious and horrible.
43:09I basically anytime I felt like I needed to describe something to Claude, instead of just doing it that one time and then having to repeat it later, I'd be like, oh, Claude, in order to do this task, I feel like you need to know about my differentiators. Maybe interview me. And I literally, Claude will interview you. You can just be like, interview me about my business, and then it will ask you a bunch of questions,
43:31and then it will write a context file for you based on what you learned. Got it. Okay. So that's a that's a lot of how I did most of this stuff. Okay. Yeah. As long as you just sat down one day and made all these folders. Right? It's not No. I add to it, like, gradually over time. So what about, like okay. Last question. Let let's say I'm just overwhelmed by all this stuff. Like, this is Yeah.
43:51Like, I wanna learn stuff, but, like, it sounds like you really personalize it for your life. Yeah. Like, what are your three tips to get started? Let's say I already use Cloud Code for, like, co coding, but I wanna use it to manage my life. The first thing is with context, you don't have to get here overnight. Like, I think the simplest rule is whenever you find yourself explaining
44:12context to Claude, stop and think about, am I ever gonna have to explain this context to Claude again? And honestly, the answer is probably yes. So instead of explaining it to Claude, change tasks just for a minute and just be like, Claude, I need to explain some context to you. Let's capture it in a context file. Got it. Yep. That makes sense. You probably need to think through, like, at a minimum,
44:33do you work with Claude at work and in personal stuff? Then maybe set up a work folder and a personal folder. That's it. Right? And then I think from there, you can just do it iteratively over time. And that's what I did. Like, I literally never sat down and just created all these files. I just and some of them are just stubs. Right? They're not that complete.
44:55But I know I'll add to them with time. And then this file we're looking at, this profile, I never wrote a word of it. Every time I have Claude add a context file, I say what index needs to be updated, and it just figures it out. Got it. Okay. So maybe, like, the process here is, like, think about your weekly calendar. What's, like, the most like, what's taking most of your time? And then and then, uh, try to you know, you have Claus, your employee, try to get it to, you know, give a bunch of instructions, backgrounds, context, have it save a bunch of files, and just work with it to start doing stuff. And then at the end of every conversation, ask it to, like, add more stuff to context or, like, improve its prompt. That's kind of
45:31that's kind of what you do. Right? Yeah. Okay. I think there's two steps to this. The first is I have worked with an admin for, like, ten years. So I'm really good at, like, looking at things that I can delegate. And for anybody who's ever delegated something to another person,
45:47what makes that work well is to have, like, a good, like, standard operating procedure for the task you're giving them. Right? It's not you can't just say to someone, another human, like, hey. Go do this thing and have them expect them to be able to do it the way that you would do it. And so if you want them to do it the way that you would do it,
46:06you have to provide for them, like, a process. Like, here's how I do it. Here's the outcome that I want. And so, like, I do this with my admin, but I record videos. I record a video of how I do things. She then looks at the video. She creates, like, a Trello checklist for herself, and then that becomes, like, our standard operating procedure. So we've been doing this for years. And so I'm already good at this muscle of like, what can I delegate?
46:30So when I started doing this with LLMs, I thought about it as like, okay, well, Claude is just a person on my team, how would I delegate to Claude? And so I started by literally looking at my own Trello board and, like, what do I need to work on today? And then for every task, I got in the habit of, like, how can Claude help with this? Got it. Got it. And then there's pieces for some tasks. Like, I don't wanna automate them. I enjoy doing them. Like, I enjoy writing. I don't want Claude to write for me. So I don't wanna automate that. So I think about, like, how can Claude augment this? What do I do while writing that Claude could accelerate? Well, research and web searches
47:05and, like, coming up with alternative analogies. And then there's other tasks like send this stupid receipt to my finance system. I don't need to ever do that. I could automate that. So then I look at, like, what do I have to do to give Claude context to be able to automate this completely?
47:23CTASo Okay. I didn't even start with the things that take the most time. I literally just forced myself all day every day every time I do a new task to think about, like, how can Claude help? Got it. Alright, man. That that that that that's a great tip to end on, I I think. Alright. Cool. Just hire hire a $100 employee and then get to help on different things. Yeah. And, I mean, it's funny that I even thought twice about upgrading from $20 to a $100. I started on the $20 plan. I only had to bump to the $100 plan when I was doing
47:56CTAboth a lot of coding in one week and a lot of writing in one week. So it's like I was literally pushing hard on Claude for a long time. So you can get pretty far on the $20 plan. Yeah. Yeah. Yeah. Yeah. And, hopefully, Anthropic guesses limits together. You know? Yeah. So yeah. Okay. So to find your detailed Cloud Code guides, just go to producttalk.org.
48:18CTARight? Is that the right site? Yeah. It's like, we can take a look right here. It's producttalk.org, and it's all, like, all my recent articles. You can see what is Claude. Stop repeating yourself. Probably by the time this episode goes out, that safety article will be live. Awesome. I'll I'll definitely read it. Yeah.
48:37CTAThere's a ton here. I'm gonna be writing a lot more. Like, a lot of my goal for the next several weeks is just to help make Claude code accessible for people, for nontechnical people, and to really start to show the power of, like, pair working with Claude. So if listeners were, like, excited about this episode, definitely check out this series. It's been a fun to write, and I'm I'm even gonna be hosting Claude Code office hours. So if you wanna even come and get help,
49:04CTAuh, I'm gonna be doing that every month. Awesome. Awesome. You you should should be there. You should be Anthropic's postpartum person. Yeah. That sounds great. Yeah. Yeah. I know. Someone asked me if I was being sponsored by Anthropic, and I'm not. Uh, if I was, I would disclose that. I really am just a big fan. Yeah. I'm I'm really a big fan too. I'm really a big fan of, like, Kat Boris and the team and what they're building. And,
49:24CTAyeah, it's funny because they're they're really kind of trying to build towards the coding use case, but, like, you know, there's so much more you can do to build a really good agent. Yeah. Cool. Alright, Theresa. Thank thanks so much for your time. Thanks for having me. This has been fun.
— full transcript
§ 05 · For Joe
Teresa built what Joe is selling.
JoeFlow product validation
Her dual-terminal morning launcher is JoeFlow's Sessions panel with a CLI instead of a UI — build the app that gives everyone Teresa's system without the months of setup.
- The `today` command format is the exact template for JoeFlow's morning batch: one trigger, multiple agents fire, one briefing file appears.
- Dual-terminal isolation (tasks terminal / writing terminal, each with its own CLAUDE.md) maps directly to Sessions panel rows — this is the architecture to show in demos.
- Process notes before compaction is the manual version of what a proper session memory layer would do automatically — a clear JoeFlow feature gap to close.
- Teresa's self-maintaining context files (ask Claude at session end what to add) could be a one-tap JoeFlow post-session prompt.
- Her 3-layer system (global prefs → project rules → reference files) validates Joe's own CLAUDE.md architecture — show this episode to anyone who asks how to set up theirs.
- Key demo moment: she writes 9K words in 1.5 days. That's the number to put in JoeFlow marketing — 'Teresa's system, out of the box.'
§ 05 · For You
One rule to get started today.
If you want Claude to stop forgetting everything
Whenever you find yourself explaining context to Claude that you'll probably have to explain again — stop and capture it in a file instead.
- The minimum viable start: create a work folder and a personal folder. Launch Claude inside whichever one is relevant to your task. That's it.
- Don't create context files upfront — just do it in the moment when you realize you're explaining something for the second time.
- At the end of every conversation: ask Claude 'what did you learn about me that should go in a context file?' Let Claude draft it, you approve.
- Keep your global CLAUDE.md short. It loads every single conversation — including when you ask Claude if your dog can eat grapes.
- When you feel like Claude is 'getting dumb' mid-session: it's not the model, it's the context window filling up. Write a summary of where you are, clear the window, paste the summary back in.
