0:03 [snorts] 0:14 We're live, Kate. 0:18 >> Thank you. Thanks. Hi. 0:21 Um, hello everyone. Thank you to Katie 0:23 for for alerting me that we're on. Um, 0:26 we are now live. Um, thank you for 0:28 joining us. My name is Kate Lee. I am 0:30 the editor and chief of Every. We are 0:32 the only subscription that you need to 0:33 stay at the edge of AI. We publish 0:36 ideas, apps, uh, and and do trainings. 0:39 And you may have come to us because you 0:41 know us for our bio checks, um, where we 0:44 get early access to new models and, um, 0:47 run them through rigorous testing uh, 0:49 and publish our findings to you, the 0:51 audience. Um I'm here today with two 0:54 members of the a of the editorial team. 0:57 Um who I will let introduce themselves 0:59 shortly, but we have done a series on 1:02 writing with AI and uh we'll share a 1:06 little bit about what we've done there. 1:07 And we wanted to cap it off with what 1:10 we're calling a write along. Um so 1:12 before we go any further, Katie, go 1:14 ahead and introduce yourself. 1:16 >> Hi, my name is Katie Parrot and I am a 1:18 staff writer here at EveryY. Um, I write 1:21 across a variety of columns that we 1:23 have. My personal column is working 1:26 overtime where I write about how AI is 1:28 changing work and then I write our vibe 1:31 checks and our context window daily 1:34 roundups 1:35 of the the latest you need to know in 1:38 AI. So that's that's me 1:41 >> and Jack. 1:43 >> Hey everyone. Uh, I'm Jack Chang. Um, 1:45 senior editor here. Um, I edit all of 1:48 the kinds of pieces that Katie writes. 1:50 um and uh write um a few of my own as 1:52 well. Um should also mention that um 1:55 Katie and I are both on uh the frontier 1:58 team here at Abri, which means that um 2:01 we're also um sort of tasked with 2:04 prioritizing our like just experimenting 2:07 with different um tools, different ways 2:10 of working. Um, I see, you know, uh, 2:12 Betty's comment in in the chat about my 2:15 post on on Jev and and just like, yeah, 2:18 trying to, you know, make it do really 2:20 interesting things. Um, so yeah, so very 2:24 excited to be here to to share more 2:26 about um, my uh, editing process. 2:29 >> Awesome. And our colleague Mike will be 2:32 joining us uh, in a little bit as well 2:34 to show his process. Um, so just to set 2:36 this up, we decided to do a series on 2:39 writing with AI because first of all, 2:41 it's something we already write about 2:43 regularly. That's something that both 2:44 Katie and Jack and Mike and everyone on 2:47 our team and Dan have been writing 2:49 about. Um, but it also felt like uh 2:52 thing the discourse had sort of reached 2:54 a little bit of a fever pitch this 2:55 summer with um lots of just different uh 2:59 news organizations and journalists and 3:01 others um sharing policies, sharing 3:04 opinions about writing with with AI. And 3:07 we felt like it was um it was just a a 3:10 good moment for us to share our what 3:12 we're doing, why we're doing it, what we 3:15 believe, and how we think you can do. um 3:18 you know your best work with AI when you 3:21 know how to use the tools well. Um this 3:24 has also grown out of uh a writing camp 3:27 and a writing workshop that Katie uh and 3:29 I have done. I think we did one or two 3:31 of them earlier this year where people 3:33 were just um incredibly curious about 3:36 the process and the actual process. And 3:39 I think one thing we um you know try to 3:42 emphasize is that like this is a process 3:44 that works individually for us. it works 3:46 for Katie or it works for Jack, but it 3:48 may, you know, you can hopefully adapt 3:50 it for yourself uh for what works for 3:53 you. Um, but we thought that we would 3:55 use this time to basically literally 3:58 take you through the process that Katie 4:00 goes through when she's writing 4:02 something, the process that Jack goes 4:04 through when he's editing something, and 4:05 the process as well that Mike goes 4:07 through when he's writing something 4:08 because his process is entirely 4:11 different from Katie's in pretty much 4:13 every way. Um, I also do just want to 4:16 show you that um if you haven't seen it 4:18 yet, 4:20 uh if you haven't seen it yet, the um we 4:22 did publish uh uh Katie's guide to 4:25 writing with uh to compound writing, 4:28 which is a plug-in that we do have in uh 4:32 available to you. Um and that 4:35 essentially allows you to uh incorporate 4:38 her methodology 4:40 uh into yours. Uh, and so it's really a 4:42 way a way to um to uh to to do that for 4:46 yourself. So I'm going to stop screen 4:49 sharing and I'm gonna go over to 4:52 Katie. Let's get started. 4:54 >> Let's get started. Um, bear with me 4:58 while I talk my way through finding my 5:02 window. Um, 5:05 >> can you share my Can you see my screen? 5:08 >> Yeah. Uh, yes. 5:10 >> Okay. Okay. So, we see Claude. 5:12 >> No, we see Streamyard. 5:14 >> Oh, we see Streamyard. Oh, that's not as 5:16 exciting. Um, 5:19 let's see. 5:21 Share screen. 5:23 Um, window Claude. Entire screen. That's 5:27 what I want. Um, 5:28 >> yes. 5:29 >> Great. So, now we see Claude, 5:33 >> right? 5:34 >> We do. Yes, we do. 5:35 >> Great. So, we're going to be writing 5:37 inside Claude. This is a little bit of a 5:40 vibe shift for me because um backstory, 5:44 Claude was actually my first love as a 5:46 writing model. [snorts] Um I remember 5:49 back in the days of Sonnet 4.5, 5:52 the team was going crazy for it. I you 5:55 know I started using it for writing fell 5:56 in love with the writing style but 5:58 somewhere around the opus 4748 6:02 definitely opus 5 era um you know opus 6:06 just stopped uh cloud models just 6:08 stopped working for me and I kind of 6:11 refugeeed over to chat pt but um in the 6:16 with the release of opus 55 I really 6:18 think that uh claude has gotten its 6:20 groove back uh shout out to the team at 6:22 at anthropic for amazing work. So, I'm 6:24 going to be working inside of the Claude 6:27 desktop app in code, which might sound 6:30 weird to people that I'm not in the the 6:33 main Claude the for artist formerly 6:35 known as Co-work. Um, but I um I I like 6:40 this is just where I live and I it makes 6:42 me happy. Um, so before I go into the 6:46 actual drafting in Claude, I do want to 6:49 just give a little bit of context about 6:51 the context that is driving this 6:53 behavior. So 6:54 >> for sure, 6:55 >> um, 6:56 >> for those that don't know, um, 6:58 [clears throat] 6:59 like folders are kind of the driving 7:01 force for a lot of us here at every we 7:03 even have a weekly session called show 7:06 us your folders at our standup where uh 7:08 on show and tell on Fridays where we 7:10 show each other our folders and how they 7:12 help us do our work. The drafting 7:14 folders that I have, I have one for each 7:15 of the kinds of columns I do are the 7:17 engine behind all of the writing that I 7:19 do. And there are really um three or 7:22 four key things to know about. The first 7:25 is my style.md. 7:27 So style defines um the column, what it 7:31 is, how it works, and what a good uh 7:36 example of the of the um of the column 7:39 looks like. So, we have the purpose 7:41 working overtime is a first-p person 7:42 ethnography of the ways AI is changing 7:44 work. Uh, it is a first-person 7:46 laboratory where Katie Parrot documents 7:48 the emotional, cultural, and practical 7:50 reality of working with AI while 7:51 actively experimenting with her herself. 7:53 And then I have reader payoff. So, a 7:56 reader should leave with a new lens or 7:58 framework that they can apply to their 7:59 work and language that can help them 8:01 better understand their own relationship 8:03 with work in the era of AI. 8:05 Oh, we we lost it. We lost it. Live 8:08 demos. Where do we go? Um, 8:11 my computer is completely freaking out. 8:14 Um, so I apologize for that. 8:17 >> Okay. 8:17 >> Um, 8:18 >> do you need to do a restart? 8:20 >> Um, I shouldn't. Let me just like let's 8:23 just chill out for a second. Um, I can 8:26 voice over a little bit about of what um 8:29 what's else is in the style MD. So that 8:32 that covers um Oh, it came back for a 8:34 second. Um, but I'm not going to click 8:36 it again. That covers structure. So, 8:38 there's something in there called the 8:39 friction to framework arc, which is 8:41 actually something that I uncovered 8:43 through the process of just feeding my 8:45 essays to Claude and asking, "What do 8:47 you notice? What do these have in 8:49 common?" And it found this pattern in my 8:51 work where I start with a personal 8:53 experience of struggle, conflict, um, or 8:56 an aha moment, and then we go from the 8:58 friction through the example to a 9:01 framework that the reader can use. So, 9:03 that's something that we're going to be 9:04 trying to work through in this example. 9:07 And then the other thing that lives in 9:10 the folder that's really important. Oh, 9:12 well here it is. Um, that's really 9:14 important. Um, oh, here are some like 9:16 the sex success equation. Um, narrative 9:21 driven analysis, personal vulnerability. 9:23 Everyone who reads Working Overtime 9:24 knows I love nothing so much as to 9:26 overshare. Strategic use of humor and 9:28 wit. And you just see that um, you know, 9:31 this is just um, just big picture 9:34 guidance. Things to watch out for 9:36 equally important in your style. MD, 9:38 argumentative weaknesses, straw men, 9:40 both sides, fallacy, flattening 9:42 binaries, here's our friend, not X but 9:45 Y. Let's get that out of there. Hedging, 9:49 and then a pre-publication checklist. 9:51 So, that's everything that's in the 9:52 style MD. And then the voice MD I'm not 9:55 going to show because it's not quite as 9:57 exciting, but it just captures sentence 9:59 level patterns. um the way I like to 10:02 architect sentences 10:04 so that you can sort of see the um how 10:07 an idea progresses. And so now that 10:10 we've kind of done that, we're just 10:12 going to run through the process in 10:13 compound engineering or compound 10:15 writing. Compound writing uh descends 10:18 from compound engineering, which is 10:19 Kieran Clawson's amazing plug-in and 10:22 framework for uh compound engineering 10:25 for for software engineering that gets 10:27 smarter and learns with you. Um and the 10:30 core engine of that system is this 10:32 pipeline of steps you go through. And 10:34 the thing about writing is that it 10:36 follows similar steps. You start with 10:38 brainstorming. You go through a sort of 10:41 planning phase which is in the in um in 10:44 writing is the outline. Um you go 10:47 through uh drafting which is the work 10:50 phase and then you kind of review and 10:52 give feedback. So, um, I have an idea 10:55 for a working overtime article. And so, 10:57 where this is going to start, uh, is I'm 11:00 literally just going to t double tap 11:02 monologue 11:04 and say, I have an idea for a working 11:07 overtime essay about how I hopelessly 11:09 messed up my context and basically 11:12 ruined my life because the models 11:14 weren't behaving anymore. Can you 11:16 interview me one question at a time to 11:18 draw out my thinking on this topic and 11:21 we'll take it from there? 11:25 Okay. So, I'm going to get rid of this 11:27 just so it doesn't really matter if you 11:29 catch a couple stray thoughts, but um 11:32 and then we're just going to let uh let 11:34 Claude take it away with some questions. 11:36 It's going to um check in. It's going to 11:39 call the skill, which I love. It's 11:40 looking at the style guide, the idea 11:42 farm, um all of these things. Um don't 11:46 get too nervous when it says a command 11:48 failed. I don't usually find that that 11:50 works. Um, yeah, this sounds exactly 11:52 like the piece in your idea form. 11:56 Um, okay. We're going to ignore the 11:58 existing 30%. Um, I'd like to draft this 12:02 from first principles because we're in a 12:04 demo. Um, can you start the interview 12:07 process over and we will um we will take 12:10 it from there. Um, but you see it it 12:13 found the ideas like all of these things 12:15 like my idea farm is where I capture um 12:19 capture ideas that I want to grow into 12:21 potential columns. Um, and then outlines 12:25 and drafts are all saved so I can come 12:27 back to things. But for the purposes of 12:30 this demo, we don't want to like like 12:32 skip the line. So, what 12:34 >> quick question, Katie. Quick question 12:36 just uh actually from from Kashik, our 12:38 colleague, who wants to know why you're 12:39 using monologue instead of Claude's 12:41 voice mode. 12:42 >> Um, honestly, it's just it's I love 12:45 Monologue. It's an amazing product. It's 12:47 in every product. Shout out to Naveen, 12:49 uh, the general the the the the 12:53 brains behind the the process, who we 12:56 actually just, uh, shared that he built 12:58 an entire language model for voice mode. 13:01 Um, so we love monologue and um and 13:06 that's just that's where we do it. 13:07 >> So it's like habit. You love it. Go at 13:09 it. 13:09 >> Yes. I I I'm very much leaning into the 13:11 familiar. This is everything is just 13:13 muscle memory for me at this point. Y 13:15 >> um and this is and this is how I work. 13:17 So what happened? Tell me about the 13:19 moment you realize the models weren't 13:20 behaving anymore. 13:22 >> I was trying to write through a draft of 13:25 an essay about um the relationship 13:28 between work and play in my AI life. 13:31 and it just went completely off the 13:33 rails. Um, I couldn't get the models to 13:35 do what I wanted. Drafts were coming out 13:38 overly crowded and flat. There wasn't 13:40 the shape that I was used to seeing. And 13:43 no matter how much feedback I gave, no 13:45 matter how much I tried to sort of brute 13:46 force my way through the drafting 13:49 process, 13:50 um, it just wasn't working. And this was 13:53 a sudden change. Um, you know, I had 13:56 oneshotted a vibe check for um for uh GP 14:00 chat GPT's uh or OpenAI's GPT6 Astra 14:05 with no trouble. And then this was 14:07 literally the next piece I tried to 14:08 draft and it had gone completely 14:10 haywire. And the thing that had happened 14:12 in the middle of that was that I had 14:15 gone a little bit crazy um and started 14:19 layering way too much stuff into my 14:21 context. I made the disastrous decision 14:24 decision to tell my machine to save 14:26 literally everything that it produced, 14:29 every outline, every decision, every 14:31 piece of feedback, um, every experiment. 14:35 And the result was just an absolute 14:38 catastrophe in terms of, uh, my contacts 14:41 ability to support me and do work the 14:43 way that I wanted to do it. So, that was 14:45 that's the origin story behind where we 14:48 find ourselves now. And the thing, the 14:52 reason that this is labeled 14:55 uh uh is that I came across this word. 14:58 I'm a German major, so I speak German, 15:00 but that's not how I I didn't magically 15:02 know this word. Uh Emily Campbell, who 15:05 is uh somebody who works at AI at Figma, 15:09 um had tweeted that she just learned 15:12 this word, which means uh literally 15:16 worsen better. So it refers to the 15:19 process of trying to make something 15:20 better only to make it worse. Uh and 15:22 that's exactly what I had done with my 15:24 context. I had um just loaded it down 15:28 with so much stuff well intention to try 15:30 to make it better that I absolutely 15:32 wound up making it worse. So that's my 15:35 going to be my first brain dump. And 15:37 you'll see it's very long. I kind of 15:39 just let myself ramble a little bit. 15:41 This is why I prefer speaking out loud 15:44 um a lot of the times, especially at 15:45 this early stage of the process. I I'm 15:49 just a a verbalizer. I'm a yapper and um 15:52 I kind of find my way to what I think 15:54 through the process of narrating. So, I 15:57 realized I could have hit enter and been 15:58 talking and had this thinking churning 16:00 while we um go. But look how fast Opus 16:03 55 is, you guys. It's amazing. Um so, 16:06 take me back to when you told the 16:08 machine to save everything. what were 16:09 you hoping it would do for you? Um, this 16:12 ties to my new role on the frontier team 16:15 where we are actively experimenting on 16:17 the frontier of AI and I just wanted a 16:21 methodology for capturing experiments 16:23 that I was doing, learnings I was 16:25 accumulating, processes that were 16:28 evolving. Um, and I just thought it 16:30 would be great to capture all of that 16:31 both for future learnings and like 16:33 retrospectives and then also for 16:35 potential writing. Um because this is 16:37 actually something that I do a lot is uh 16:40 in when I'm trying to reconstruct the 16:42 lessons I've learned for working 16:43 overtime piece for example, I will very 16:45 often go back into my saved context, my 16:48 save the chat histories and things and 16:51 to reconstruct what I did. Um so this is 16:53 something that I've been able to rely on 16:55 in the past, but I just overdepended on 16:58 it. Um and um 17:03 and it turned out to be a horrible 17:05 mistake. 17:11 You said this was something you'd been 17:12 able to rely on before. What was 17:14 different this time? Where where did 17:16 ordinary reliance turn into 17:17 overdependence? Okay. And this kind of 17:20 highlights something sometime, 17:21 especially uh Claude models, they will 17:24 get a little they they are very 17:25 psychologically insightful and they will 17:28 ask you tough questions that actually 17:30 make you um wrestle with um the with 17:34 what you're thinking and and the 17:36 consequences of it and sometimes take it 17:38 in a direction that you weren't 17:39 necessarily thinking about. So for 17:41 example, where did ordinary reliance 17:43 turn into overdependence? Like I don't 17:44 know if that's necessarily 17:47 what I want to drill into in this piece. 17:50 Um I'm more because like that's a 17:52 separate question of over depending on 17:54 your context versus just having context 17:56 you can depend on. 17:58 >> Um so I'm going to say that um I I'm not 18:01 necessarily as interested in exploring 18:03 overdep the idea of overdependence as 18:05 just as much as what happens when 18:08 context that you've historically been 18:09 able to trust kind of betrays you. 18:16 And we should state Katie that you were 18:18 this was earlier this week. You were 18:20 working what you were working on was the 18:23 very vibe check for Opus 55 that um 18:25 hopefully many of you have read along 18:27 with um some inputs you gave into the 18:30 GPT soul. I'm not even remembering which 18:34 number it was but uh which just came out 18:36 on Tuesday as well. So um this was 18:38 literally three or four days ago. 18:41 >> Yeah. How did you figure out the context 18:43 was the problem? What was it like to 18:45 realize the thing you trusted was the 18:47 thing going wrong? I kind of covered 18:50 this already, but it really was the only 18:52 variable that had changed in between the 18:55 successful Astrovibe check and this 18:58 workplay piece that had gone completely 19:00 haywire. Um, so it was about um, yeah, 19:04 just knowing that's what changed. I 19:06 didn't change. The models didn't change. 19:09 They've changed since, but they hadn't 19:11 changed at this point. Um, and honestly, 19:13 what was it like to realize the thing 19:15 you trusted was the thing going wrong? 19:16 Uh, it was terrifying. Um, it was 19:20 really, really worrying because work 19:24 doesn't stop. You need to keep making 19:25 progress through your commitments and 19:27 realizing that this essential 19:29 infrastructure that I was really relying 19:31 on had gone haywire. 19:33 um you know I needed to stop what I 19:36 needed to do was stop everything and 19:39 rebuild from scratch which I actually 19:40 have done and we'll talk about that when 19:42 we get to solutions but um at the time I 19:46 was trying to throw oneoff fixes into 19:50 the context I was and I was actually 19:53 making it worse by layering on more 19:55 instructions and more contradictory 19:58 um inputs that just made the models more 20:01 confused because I was just layering on 20:03 more and more and more into these 20:05 documents like my style.md and my 20:08 voice.md and my agents.mmd 20:11 and the whole thing just got hope 20:13 hopelessly tangled 20:16 lessly 20:17 tangled 20:19 I cut it off a little early there um 20:23 it's okay when you finally looked inside 20:25 style.md and and yada yada what did the 20:28 tangle actually look like is there a 20:30 specific instruction or contrad 20:32 prediction that sticks with you. And 20:34 this is honestly where I would probably 20:36 send the model looking like exact. This 20:38 is exactly the kind of thing that I have 20:40 on my machine. It's not on this machine 20:42 though. It's on my Mac Mini. Um, which 20:44 is a journey I'm currently on is trying 20:46 to harmonize the documents available on 20:48 both of my machines. Um, but uh it's 20:53 less it's less specific. But what I 20:55 noticed in my working overtime context, 20:57 for example, is that there were like 20:59 five different templates that I had 21:01 meant to be different templates that the 21:04 model could choose from that had somehow 21:06 gotten fused together into like an an 21:10 impossible list of conditions that every 21:12 draft had to meet, you know? So, it had 21:14 to have a story and a cultural 21:17 connection and a an actionable framework 21:20 and funny joke and like it just it just 21:23 it completely broke the model's ability 21:25 to help me decide which of those 21:28 building blocks made sense for the piece 21:31 at hand. 21:39 Okay. Who do you picture reading this 21:41 piece? what might they be doing in their 21:43 with their own context right now that 21:44 makes the story matter to them? 21:47 I think context is really um something 21:52 that 21:54 you know you kind of come to the 21:55 realization of how important it is to 21:57 how you're building um the the farther 22:00 you get into working with AI. So for 22:01 example, I was talking to a friend of 22:03 mine who's a little bit farther behind 22:05 me in the sort of AI journey and I was 22:07 telling him how I had messed up my 22:09 context and he was like wait what is 22:11 context why is it important and then I 22:13 explained to him kind of a brief history 22:16 of the concept of context engineering 22:18 which started in the developer space and 22:21 refers to this orchestration of 22:24 instructions preferences um style guides 22:27 that you bring together to allow to make 22:31 AI behave the way that you want it to 22:33 behave. Um, and so this is a a a topic 22:35 that I kind of caught on to around the 22:38 time that projects became a thing and 22:40 claude. Um, however many eons ago that 22:42 was in AI time. Um, I started I I I I 22:47 started adding examples to projects. 22:49 Then I I created these style guides 22:51 again just based on my own like and 22:54 bottomless curiosity about myself and my 22:57 writing process. Um, and the output and 23:01 I found very quickly that the outputs 23:04 helped shape pieces 23:07 like the way that I wanted them to be 23:09 shaped. And then when I made the 23:11 transition over into the the wild west 23:14 of Claude Code and Codeex GPT 23:19 um so much more became possible with 23:21 having you know these files on your on 23:23 your desktop that are interrelated that 23:25 you can orchestrate together to um and 23:28 combine with a plug-in like compound 23:30 writing to get the AI to work the way 23:32 that you want to to work with it. So, I 23:34 think where this becomes interesting to 23:36 our readers, um, you know, maybe they 23:39 don't have context yet, and this is an 23:41 early cautionary tale to them to keep it 23:43 simple, stupid, um, the way that I 23:46 should have, or maybe they're a little 23:48 farther along, and they've noticed some 23:49 of these challenges with, um, unruly 23:52 context and things getting out of 23:53 control. and then they can learn from 23:56 the you know the process of 23:57 disentangling and some of the principles 24:00 that I think I've landed on as far as uh 24:03 how to keep this from happening. 24:06 So, just if you're just joining us, um 24:09 just uh for those of you who have just 24:11 come on, we are going through uh this is 24:13 a write along hosted by every we are 24:15 going through Katie Parrot's uh compound 24:18 writing process to actually show 24:21 literally how the sausage is made, how 24:23 she writes a piece using her AI tools 24:27 and systems. um she has been going 24:29 through uh having her AI interview her 24:32 um to draw out what she's been uh 24:34 thinking about with this uh in this 24:36 piece. So um Katie, keep going. 24:40 So the next question is tell me about 24:42 the rebuild. What did you actually do 24:44 and which principles came out of it? 24:46 Kieran comes into play here again 24:48 because I actually just asked my codeex 24:50 like what's going on with my context and 24:53 it without me even specifying called a 24:55 compound engineering skill I didn't know 24:57 existed called doc review and it 25:01 actually went through my context and 25:02 analyzed it against five different core 25:05 principles like feasibility um product 25:08 experience three others that I have 25:10 amnesia about and it delivered [snorts] 25:12 a report to me about um what was wrong 25:16 with my context and that's where I 25:18 discovered okay all these instructions 25:20 and agents MD and style MD are 25:23 conflicting and fighting each other and 25:25 confusion and confusing each other and 25:28 so um what I did honestly just earlier 25:31 this week is I nuked the entire thing uh 25:35 I moved all of my folders into a 25:38 historical folder and rebuilt them from 25:40 scratch. Um, so I kind of went through 25:42 the compound engine, the compound 25:44 writing setup process again for my 25:46 writing folders. I reset the examples 25:50 that uh, AI had access to. I rebuilt my 25:53 style MD and my voice MD and just kind 25:56 of reset everything in a in a Marie 25:59 Condo kind of way. Let go of what was 26:01 not bringing me joy. Um, and the 26:05 principles that came out of it. Um 26:08 there's a there's a there's a couple and 26:09 I I remember them because you know I 26:11 cheated a little and went through this 26:12 process before but um you can make 26:15 something you can you know good 26:16 intentions can go bad. You can mean to 26:19 make something better and not make it 26:21 better. Um more memory is not always a 26:24 good thing. Sometimes it pays to let 26:26 your model forget things that no longer 26:27 serve you. Um prune your context like a 26:31 bonsai tree. You're kind of always in 26:33 the process of taking things away. 26:36 shaping things to the way that you want 26:38 them to be shaped. Um, so that's kind of 26:40 some of the examples of the takeaways. 26:44 Um, and then just because we're a little 26:45 short on time and I want to make sure 26:47 other people have time, I'm just going 26:48 to run through how we go from this to 26:50 the outline, a couple reviewers quickly. 26:53 Um, and then 26:54 >> one note, Katie, we do have some time. 26:57 Mike uh Mike, just for the audience, 26:59 Mike is uh at a conference and unable to 27:01 find a quiet spot to uh to stream from. 27:05 So, so it'll be the three of us. 27:07 >> It's the Katie and Jack show. Yay. Okay, 27:10 cool. So, we'll just keep rocking then. 27:12 Um, so since the rebuild, what has it 27:15 been like to work with the models? Has 27:17 anything come back or stayed broken? 27:19 >> Um, so far we're rocking and rolling cuz 27:23 we're we're going through this process 27:25 and it's going smoothly. Um, I've 27:28 written some pieces for my personal 27:30 newsletter that have gone well. I um I 27:33 test r test test drove the um the 27:36 process for vibe check through a uh a 27:39 model that we were you know that we had 27:42 not reviewed um and it went smoothly. So 27:46 I'm kind of ready to trust my context 27:48 again. That trust had been broken. The 27:50 trust in me myself had honestly been 27:52 kind of broken. Um and that's another 27:55 thing that I want to say. This is 27:56 another important lesson that is 27:58 important. the never edit your context 28:02 under pressure or duress. Um, so at the 28:05 time that I was doing this, uh, I was 28:09 kind of in the midst of like AI 28:12 psychosis. Astra had come out. I felt 28:15 like I was flying and could do anything. 28:17 And that was really the origin of all of 28:18 these amazing ideas that I thought were 28:20 so amazing to change my context. And 28:23 then, you know, deadlines happened, vibe 28:25 checks had to go out and I was trying to 28:28 fix it while I was working. And that's 28:29 just not a good environment to do these 28:31 things in. So, um, finding a quiet 28:35 state of mind from which to work with 28:37 your context is also important. 28:40 And so, we're we're getting toward the 28:42 end. I can tell from the questions that 28:44 it's asking, like, you know, who what 28:46 you want the audience to take away? Um, 28:49 what did you learn? And it says, I think 28:51 there's enough here for an outline. Um, 28:54 how do you want to go from here? 28:55 Reflect. Generate possibilities. Keep 28:58 asking. So, I'm going to have it reflect 29:00 back what it knows. Let's reflect back 29:03 what I've said and uh then you can tell 29:06 me and then we'll we'll we'll shape it 29:08 from there. 29:10 I'd say Katie, you've done this enough 29:12 times that you see you have a sense of 29:14 the questions and the back and forth and 29:16 when it's going to say essentially I've 29:18 gotten what you need or I've gotten what 29:20 I need, excuse me. 29:22 >> Yeah. And like the the thing that I'll 29:24 say is like you're kind of in charge 29:25 like you're you're very much in the 29:27 driver's seat. I think a failure mode of 29:29 this is that you give the the the model 29:31 too much agency and too much control 29:34 over things. And so there are times when 29:36 it's like again like I said it'll ask me 29:38 a question that I'm not interested in. 29:39 I'll be like let's not talk about that 29:41 or it will or I'll just get to the point 29:44 where I feel like I know what the piece 29:46 is about and I'll cut it off. But a lot 29:48 of the time and this is tied to the fact 29:51 that it has this context about the the 29:54 structure that my pieces often follow. 29:56 it naturally. It kind of pulls out the 29:58 information that it needs to to find the 30:02 the friction, find the um the action 30:04 steps, find the lessons, and then we can 30:07 move on to drafting. So, um what 30:10 happened? You on one shoted of with no 30:13 trouble. the very next piece, an essay 30:14 about your life went completely off the 30:16 rails, why you saved everything, what 30:19 you want to explore, how you knew, what 30:22 the tangle looks like. And now, 30:23 obviously, 30:25 um 30:27 I'm not going to stop and read all of 30:29 this right now because that would make 30:31 for bad TV, but um but I would stop and 30:36 um and review all of this. And this is 30:40 something that I always have it do is 30:42 open loops that I haven't resolved. Um, 30:45 your original pitch said this basically 30:47 ruined my life. You haven't yet said how 30:49 much [snorts] of that personal toll 30:50 belongs in the piece. Um, 30:55 that's that that's where like my chronic 30:57 oversharing, pathological oversharing 30:59 comes in. Um, so that is probably 31:01 something that I would do, but um I can 31:04 always add 31:04 >> You mean you would go back and specify 31:06 is what you mean? Yeah, I would I would 31:08 um 31:10 you know how much of the personal toll 31:12 like we can we can be honest about the 31:15 personal toll and the fact that I was 31:16 like already not doing so great and 31:19 getting a and I was a little underwater 31:21 and then the fact that um the fact that 31:24 my context wasn't helping me anymore put 31:26 me that much farther behind. Uh and I 31:29 you know I I'm always open to share that 31:32 that kind of stuff. I've written about 31:34 my mental health on the channel on the 31:36 column before. Um, so that again, so 31:40 that's that's that's fair game to a 31:42 certain extent. Um, Kieran comes into 31:45 play here again. Yes, it did. Oh, wait, 31:47 no, that was on the call before I 31:49 started interviewing. Um, so Kieran, so 31:52 we have Yeah, we have to establish the 31:53 context that Kieran built something 31:55 called compound engineering, which has 31:58 tools that can be applied to 31:59 non-engineering work like doc review. 32:03 Um, and then the three do review things. 32:05 I don't remember which we're not going 32:07 to dwell on. What's wrong or missing? 32:09 Um, I think we're ready at this point. 32:12 Honestly, I think I'd rather see a 10%. 32:15 Um, 32:17 so, 32:19 so a 10%. This comes from my background 32:22 in a content marketing agency called 32:24 animals. um our the founder sort of 32:27 created instituted this principle of the 32:29 10% frame outline and the 30% outline 32:32 and the idea is there are certain 32:34 problems with a piece of writing that if 32:36 you wait until later in the process they 32:39 become harder to untangle. So if you 32:41 wait till you have a whole blog post and 32:43 the framing is completely wrong or the 32:46 the reader takeaway is what are you 32:48 doing is completely off then um the 32:52 whole thing falls apart and it's really 32:54 better to fix those issues upstream. And 32:57 so I took that concept of the 10% and 32:59 30% which is like this is a 10% which is 33:02 just the story beats and the the working 33:05 thesis. Um, and so I review this, make 33:08 sure all of this makes sense, and then I 33:10 move it to a 30%. So, I'm going to look 33:12 at the thesis here because that's really 33:14 the important part. Um, context you've 33:16 learned to trust can betray you, and it 33:18 happens through good intentions. The 33:20 effort to make it better is what makes 33:21 it worse. The way back is to prune your 33:24 context, not add more to it. So, that's 33:26 that's a good place to start. Um, you 33:29 know, I might um 33:33 I might do some things. So like I 33:36 actually like have something in mind for 33:38 this a little bit. Um so I want to start 33:40 with the concept of fimbasong. So I'm 33:43 just going to go ahead and tell it 33:45 >> pronunciation by the way. But go on. 33:47 >> Yeah. German. I I I got to use the the I 33:50 got to use the the um 33:54 that the um the I got to use the the 33:57 degree that I got somehow. It was very 34:00 expensive. But so like something else 34:02 that I will often do with these pieces 34:03 is I will at the beginning of the 34:05 process rather than just the brain dump 34:08 like I I have existing notes. So here 34:11 I've got all of this stuff um that I had 34:14 written out and this is what I initially 34:16 gave the model when I did this pre-bake 34:18 that Jack is going to edit. Um so I'm 34:22 going to just give 34:24 it this part. That's the opening. Um, 34:28 I think I have an opening or a direction 34:30 for the opening that I'm really happy 34:32 with. So, I'm just going to give you 34:33 that and then we can rework the we can 34:36 build the 30% based on based on this 34:39 hook. 34:40 And a hook, for those who don't know, is 34:42 the the opening of the piece. And it's 34:44 called a hook because it needs to pull 34:46 you in and reel you in like a fish. Um, 34:49 so what I'm always looking for in a hook 34:51 is something that opens a curiosity gap. 34:54 So, it needs to make the reader wonder, 34:57 oh, what does that mean? Um, what 34:59 happened and like how did you deal with 35:02 it? Um, and so a lot of times the hook 35:06 or the intro will just come to me fully 35:08 formed or very close to it and I'll 35:10 start there. Um, so you don't h that's 35:12 another point is you don't have to use 35:14 AI for every part of the process. if it 35:16 doesn't serve you or if if you like want 35:19 to like get in there and and tra if you 35:22 will, you can totally do that. Um, so 35:26 here we have our 30% outline for Schlim 35:28 Besselong. Um, and the the the structure 35:31 here is very um very structured and this 35:34 is actually something I really 35:34 appreciate about Opus 5.5 is that it 35:37 follows this structure because I've had 35:40 some very we've had some very 35:41 know-it-all models lately who don't like 35:43 to be told how to format things. Um, and 35:46 that makes it really difficult for me to 35:48 see the information that I need to see 35:50 in order to confirm. Yes, this is the 35:52 direction that I want. So, like here we 35:54 see the the main point um name what you 35:57 did to your context with the word and 35:59 establish what the thing you broke is 36:00 how you stay sane. um the beats 36:04 um the thesis which you know I might 36:06 workshop a little bit um the promise and 36:10 then what's in the folders 36:13 uh worse and bettering round one um the 36:16 betrayal worse and bettering round two 36:20 um diagnosis and rebuild after 36:23 conclusion 36:25 um and then kicker's not set yet that's 36:28 the conclusion and then open loops so 36:31 this is Another checkpoint where if I 36:34 wasn't, you know, live, I would spend 36:36 some time reviewing this output, 36:39 changing things. You know, maybe I want 36:42 to move some things around. Maybe I 36:44 think it's a little getting a little too 36:45 long to happily fit inside 1500 words, 36:48 which is very often a problem of mine, 36:50 so I'll cut some things. Um, 36:54 uh, you know, it says the crash here. 36:56 The crash comes after the explainer and 36:58 the save everything, which is 36:59 chronological. Do you want a quick flash 37:01 of the crash earlier so the readers feel 37:03 the stakes before the backstory? Like 37:04 that is actually something that I would 37:06 say yes to. Like I want to show the the 37:08 stakes and the consequences. Um so 37:11 that's something that I wish I had just 37:13 hit monologue and told it that because 37:16 that would be efficient. But so this is 37:18 what the outline stage and this is the 37:20 point in the process where I start 37:23 bringing in reviewers. So reviewers are 37:26 my favorite part of the compound writing 37:28 plugin. They are a group of a bunch of 37:30 personas that I have built um based on 37:33 different writing principles um that 37:38 um come from some of them are tied to 37:41 specific writers and storytellers I 37:42 admire. Some of them are just principles 37:45 of good writing. So um I think what I'd 37:48 like to run here first of all is reader. 37:50 Um so reader 37:53 is 37:55 reader is a firsttime reader. This is a 37:58 cold reader who doesn't know the have 38:01 the context um on the piece and I just 38:04 want to check like is there anywhere in 38:06 this outline even that the reader might 38:08 get lost where we need to explain things 38:10 a little bit more. Um and so I have this 38:13 skill set up with this definition of 38:16 this is what I want you to look at. I 38:19 think it's things like where do you get 38:20 lost? What's disappointing? Like what's 38:23 what's missing that you would that you 38:24 would want to see? Oh, what? What? Oh, 38:27 it's not here. So, let's do Hitchcock 38:28 instead. Uh, 38:29 >> I was gonna say, can you explain, Katie, 38:31 you've got names for some of your 38:32 readers, like Hitchcock, Sorcin, 38:35 >> um, I don't know if you've Sedaris. Um, 38:38 and essentially what this panel of 38:40 readers is. 38:42 >> Yeah, it's it's so we've got Hitchcock 38:44 is for suspense. So, that's really like 38:46 the lean in. The principle that that's 38:49 built on is there's a there's a saying 38:51 that Hitchcock has about a bomb under 38:53 the table where if you have there's a 38:55 bomb under the table in your scene and 38:57 it blows up, okay, you surprise people 38:59 for one second. Um, but if you show them 39:02 the bomb and say this bomb will go off 39:04 in 10 minutes, you have 10 whole minutes 39:06 to sort of draw that out and the reader 39:08 leans in. So that's the principle behind 39:10 Hitchcock. um the principle behind 39:13 Sorcin that's inspired by the famous 39:15 walk and talk in um in West Wing in 39:19 general. Sorcin's dialogue moves very 39:21 quickly. Um and I want to I want to see 39:24 if the piece is moving along or if it's 39:26 getting slowed down anywhere. Um and 39:29 then um who else? Hitchcock is just for 39:32 concision. Um keeping it keeping it 39:34 tight, killing our darlings. Um and mom 39:38 is one that I love quite a bit. mom as 39:41 somebody who cares about you very much 39:43 but is confused by everything you're 39:45 saying. [laughter] That's 39:47 >> that's a person reader. 39:50 >> Kitty Kitty, how do you how do you 39:51 decide like which personas are you're 39:53 going to throw at a piece? 39:54 >> Um, at this point I have a kind of a 39:57 consistent set like reader I always run 39:59 because I always want to see where are 40:01 people getting lost. Um, Hitchcock I 40:04 very often run when it's a piece that 40:06 has a story. Um because I want to know 40:09 that this I want to feel like I clearly 40:11 understand. Okay, are there stakes? Um 40:13 oh, Vonagget is another one that I 40:15 really [snorts] love. That's like 40:16 Vonagget has eight characteristics of 40:19 story. Um things like start close to the 40:21 end, respect the reader's time, give 40:23 them someone to root for. So that's 40:25 another one that I would run. I don't 40:27 run Vonagget on a vibe check. Um but I 40:29 run Vonagget very frequently on working 40:31 overtime pieces that are um that are 40:33 narrative driven. Um, so that's some of 40:37 them. And then there's things like 40:38 objections I will very often run, which 40:41 is just what could somebody say no to? 40:44 Um, there's also a meaner version of 40:46 that, which I've recently renamed from 40:48 to nemesis. So that's imagining 40:51 your nemesis reading your piece. But, 40:53 um, so here we have the reader report, 40:55 which it actually found despite saying 40:57 it didn't exist. Um, so the first-time 41:00 reader, it's reading as a working 41:01 overtime reader. That's a knowledge 41:03 worker using AI. Uh, you know, anywhere 41:07 from my friend's what is context level 41:08 to someone with their own sprawling 41:10 setup. What I think the piece is saying, 41:12 the writer tried to improve the files 41:14 make her AI tools work, broke them, and 41:15 learned that pruning beats piling on. 41:18 Overall first impression, 41:21 the the reader waits a long time to see 41:23 the damage. So maybe we should have 41:26 taken that um taken that that guidance 41:29 to move that damage up the way that the 41:32 outline step had actually opened up. Um, 41:35 the reading experience, 41:37 Shod and Frea. Shod and Frea is one the 41:40 f my favorites that are in the draft 41:42 that Jack will review is um, uh, 41:47 Baken Gaz, which means a face worthy of 41:50 slapping. 41:52 And um, [laughter] 41:54 and there's another one that's like uh, 41:57 it involves bacon somehow. So, we'll see 41:59 that [laughter] 42:00 used in a sentence. 42:04 which means I have ruined my uh context. 42:09 Um I have worse and bettered my context. 42:11 This causes a stumble. Um causes a 42:13 stumble causes a stumble. So we're we're 42:17 seeing here that we're kind of like 42:20 noticing where things are lost or 42:22 missing. Um what may [snorts] put the 42:25 reader off? AI psychosis if it isn't 42:26 calibrated. That's a good call out. that 42:28 is a very specific experience that is 42:31 topical sensitive. So I might not want 42:34 to use that language in the finished 42:35 piece. Um what works on first contact 42:38 fix first the question that I still 42:41 have. So that's a reader and if I 42:44 thought that it was going to work I 42:45 would have stopped and said okay we need 42:47 to make this this and this change but I 42:50 went ahead and did another one. So, 42:52 suspense analysis, 42:54 current tension, medium. The hook plants 42:57 the bomb well, but the outline then 42:58 diffuses it for two sections. Um, the 43:02 bomb under the table is the instruction 43:03 to say everything. Um, your hook already 43:07 told readers your worst and bad you 43:09 worse and bettered your context, so they 43:10 know the disaster is coming. [snorts] 43:12 Um, right now the outline covers that 43:14 scene quickly instead of letting it 43:16 tick. So, um there there's just more 43:19 cert more suggestions 43:21 and like I've got to sit and wrestle 43:23 with these and be like um so here's a 43:26 suggested revision of sequence um the 43:28 hook um the the peak 43:32 um a bomb gets planted the explosion 43:34 round two. So like um and what changed 43:38 um where this disagrees with the reader 43:40 report. That's interesting. Um, 43:43 showing the crash earlier would turn 43:45 suspense into surprise. Reordering 43:47 around the astrop peak fixes the 43:48 momentum problem without spoiling the 43:50 fall. That's your call. So, again, this 43:52 gets really it gets really deep a lot of 43:55 the time and like obviously this is an 43:57 AI's reconstruction 43:59 of these kinds of theories of different 44:02 theories of mind so to speak. Um, but 44:05 it's so it's really it's really a 44:07 mechanism for self-reflection and 44:09 saying, "Do I agree with this? Is this 44:12 something that I would not have caught 44:14 that I that I want to incorporate into 44:16 my outline at this stage?" Um, and so 44:20 that's the kind of thing that I would do 44:23 um at this stage in the process. But for 44:25 the purposes of this session, I think 44:29 we're going to go ahead and move forward 44:31 to drafting. Uh well, there's one more 44:34 thing. Okay, we're gonna draft, but um 44:37 we're gonna go section by section. So, 44:39 this is a this is a point in the process 44:42 where I have to fight my own impulses 44:44 because I want to go fast. You know, I 44:46 want to be I want to I want to think 44:48 this thing knows me so well. It has so 44:50 much context. I've talked it through the 44:52 whole story. Surely, it can write it can 44:55 oneshot a piece. Um, and you will 44:57 actually see when Jack edits what 45:00 happens when I try to oneshot things 45:02 because it's not great. [snorts] Um, so 45:03 the the process that I try to follow 45:05 when I'm behaving myself uh is section 45:08 by section. So I'll start with the 45:10 intro. I'll often spend a lot of time on 45:12 the intro because that I find that 45:14 really sets up the whole thing. And then 45:16 I'll go section by section because even 45:18 with an outline that I've approved 45:20 through the writing process, you 45:22 discover things that change. Um, and 45:24 this is true of TRAD writing too. In 45:26 fact, many of people the objection they 45:28 have to the concept of writing with AI 45:30 is that what that the piece emerges from 45:32 the drafting and they sort of and and 45:35 they argue that um writing with AI kind 45:37 of robs you of that process of 45:39 discovering the piece through the 45:40 process. But this is my version. Um, I I 45:43 still find that it's just that 45:45 oftentimes instead of having to write 45:49 the sentence out myself and decide that 45:50 I hate it. Um, I um I have AI write out 45:55 the first draft and then that's 45:56 something for me to react to and be 45:58 like, okay, is this right or is this not 46:00 quite right? So, I'm going to say, let's 46:03 go ahead and start drafting just the 46:05 intro. Make sure we have the hook, the 46:08 bridge, the thesis, and the promise. 46:12 um which that's that those those kinds 46:14 of for people not familiar with kind of 46:18 writery language. We've talked about the 46:20 hook. So that's what pulls the reader 46:22 in. The bridge is kind of my shorthand 46:25 for the larger for zooming out to like 46:28 the stakes and why the reader should 46:30 care and kind of creating that 46:32 connection uh from the hook into the 46:35 thesis. The thesis is your main 46:36 argument. Uh, and the promise is what 46:39 the reader stands to gain from reading 46:41 the piece. It's like a a a g a gesture 46:44 toward what the piece contains to give 46:46 the reader a reason to keep reading. So, 46:48 we're going to see how our model friend 46:50 does at producing the draft. It's 46:53 calling the skill. It's reading the 46:55 voice guide. Um, it may have read some 46:58 examples in the background. I'm not 100% 47:00 sure. And hopefully it'll come out with 47:03 my opening line, which I'm actually very 47:05 fond of. Um, but we're still thinking. 47:09 >> And about how long does it take then to 47:11 to oneshot what I mean and this is a new 47:13 model so it may be quite fast but 47:16 >> yeah it's uh to oneshot a whole piece 47:19 end to end. Honest like I can tell you 47:21 the Astro the Astro vibe check took 47:23 eight minutes. 47:25 >> Wow. 47:26 >> Yeah. And like I just kind of set it up 47:28 and was like make a goal. And I probably 47:30 didn't need to tell it to make a goal at 47:31 this point. Um but um and like honestly 47:35 I'm a little surprised that it's 47:36 thinking so much at this stage because 47:38 um I haven't really seen it take this 47:41 long, but sometimes it does. And this is 47:43 the point where Hello. Here we go. Um so 47:47 we're just going to read a little bit. 47:48 So it's a wellknown fact of the universe 47:50 that the Germans have a word for 47:52 everything. Commerc literally grease 47:55 bacon is the run from emotional eating. 47:57 Bacon is there to panic. Gate closing 48:01 panic is the fear that time is running 48:03 out to do the thing you're supposed to 48:04 do with your life. 48:05 >> Oh my god, that's amazing. [laughter] 48:08 >> Recently, this gem came across my ex 48:10 feed for literally worse and bettering. 48:13 It refers to the most human of impulses, 48:16 our desire to make something better only 48:17 to wind up making it worse. Used in a 48:20 sentence 48:23 or I have worsened bettered my context 48:26 reader. That's precisely what I have 48:27 done onto the folders on my desktop, 48:29 which is inconvenient because those 48:31 folders are the main way I stay sane out 48:33 here on the decidedly unpaced AI here. 48:36 Uh, inside them is what people who build 48:38 with AI call context. The style guides 48:40 and samples and instructions that tell 48:42 and chatb2 who I am, how I write, and 48:44 what I'm working on. When my context is 48:46 good, the models hand me drafts that 48:48 already have the shape I'm after. Then 48:51 mine went bad. And like this is this is 48:53 an example of something I might change 48:54 because like that when my context is 48:56 good I expect and I think this is an 48:59 expectation that I would want to lean 49:01 into is like I want when my context is 49:03 bad to come after that. Um so this is 49:06 just you know and then and like I think 49:08 the the the the model is probably trying 49:09 to avoid repetition there because that's 49:11 something that usually I'm looking out 49:13 for because models will often state and 49:15 restate the same idea over and over 49:17 again. So I think it's it has a good 49:19 impulse here but this is something I 49:21 might steer I might push back on and be 49:22 like let's construct it as um when my 49:26 context is bad yada yada. So drafts come 49:28 back crowded and flat feedback bounced 49:30 off them. I was already a little 49:32 underwater and the system I'd built to 49:33 keep me afloat started pulling me down 49:35 with it. For [clears throat] a while I 49:36 stopped trusting my context. Then I 49:38 stopped trusting myself. 49:40 >> The worst part is that I had done this 49:41 to myself on purpose with a clear 49:43 conscience and a [laughter] lot of 49:44 enthusiasm. 49:46 I gave my context more memory, more 49:48 rules, more of everything. And each 49:49 addition made it a little worse at the 49:51 job I built it to do. Context you've 49:53 learned to trust can turn on. And good 49:56 intentions are how it happens. The way 49:57 back for me was subtraction. If you 49:59 haven't built context of your own yet, 50:01 consider this a cautionary tale. Keep it 50:03 simple, stupid. 50:05 [laughter] Um, I don't know that I that 50:07 was more like on the fly writing with 50:09 which sometimes I'll come up with some 50:11 good lines and I get happy when they get 50:12 included, but I don't need no keep it 50:14 simple stupid, especially not. Um, if 50:17 you have and you're you watch yours get 50:19 unruly, I'll show you how I figured out 50:21 what went wrong, how I tore the whole 50:23 thing down, rebuilt it, and the 50:24 principles I'm using now to keep it from 50:26 happening again. Um, 50:29 honestly, pretty good, man. Um, 50:33 >> like this is it's like this you I was 50:36 able to read it really naturally because 50:37 this sounds like me to me. Um, and 50:41 that's that's really what I that's all 50:43 you can ask for in a in a in a first 50:45 draft is something that's close enough 50:47 that yeah, you want to get in there and 50:49 tweak here and there. Um me and like um 50:53 you know I in a piece I wrote called uh 50:56 keep your how to keep your writing weird 50:58 in the age of AI. I talk about the need 51:00 to kind of get it up and get in there 51:01 and rough it up and like take some of 51:03 the smooth pros and make it weirder. But 51:07 that was in an era before I had this 51:10 whole comprehensive system that does a 51:13 lot of that roughing up for me because 51:15 it uses the language that I used in the 51:17 interview and it has examples and it has 51:20 guidance about my my sense of humor uh 51:23 my use of examples and all of that kind 51:25 of stuff. So things to check uh the 51:28 German examples both are real. That's 51:31 good. Typoix, thank you. I have is not 51:35 right. Um self trust. I really like that 51:38 quite a bit. Thesis wording is pretty 51:41 good. Um the way back from me was 51:43 subtraction. We do need to credit Emily 51:46 Campbell um because I don't want it to 51:49 seem like I um came across this on my 51:53 own. Um, and then it says, "I haven't 51:56 saved this anywhere. Tell me if you want 51:57 it in drafts or we can keep working on 52:00 it here." And that's something that I 52:02 instituted [laughter] as a lesson from 52:04 this whole uh disaster that we're sort 52:07 of reconstructing is I don't want every 52:10 single outline saved. you know, I want 52:13 certain checkpoints saved so I can go 52:15 back, re-examine choices, um, pull 52:18 things back in if I like, you know, if I 52:21 draft something and then draft it 52:22 different and then decide I want to 52:23 revert back to the old thing. Version 52:25 control, as we say, as we call it, um, 52:28 gets a little tricky when you're working 52:30 on your desktop versus, you know, in a 52:33 Google doc where it just naturally saves 52:35 all that stuff. But, um, so I I try to 52:37 be deliberate about what I save and 52:38 don't save. So, I'm gonna say let's go 52:42 ahead and save this intro. Uh, I'm quite 52:45 pleased with it. Um, 52:48 yeah, I think we can go on and draft 52:50 section one. And this almost never 52:52 happens, by the way. Um, you usually I 52:55 have to wrestle with it to get the 52:56 thesis right. I have a thesis skill that 52:59 gives you three different constructions 53:00 of the thesis, so you can pick which one 53:03 you agree with. You you go back and 53:05 forth to get the argument right. But um 53:07 for whatever reason, the the writing the 53:09 the AI writing gods are with us today 53:12 probably because we're with we're on OBS 53:14 uh 5.5 which is just a a lovely model to 53:18 work with. [snorts] Um and um it's going 53:20 to think because it's got to save the MD 53:23 file to the folder. Um 53:25 >> where are Let me just do a quick 53:26 interruption and just say where we're at 53:28 because we are now 55 minutes into this 53:30 live stream and Katie, you've taken us 53:32 through 53:33 >> um quite extensively your process. Um 53:36 where we are now is um Katie has taken 53:40 through the compound writing process uh 53:44 her AI in this case it's opus 55 has 53:47 interviewed her 53:48 >> um based on an idea that that she 53:51 initially had um created two outlines 53:54 one a 10% outline one a 30% outline that 53:58 um Katie checked both of those Katie 54:01 also had her um her reviewers uh her 54:04 reviewers ers uh her panel of viewers 54:06 which are based on um personas like 54:09 Hitchcock uh Sorcin Vonagget um review 54:13 the piece for and the outline for 54:15 certain for certain qualities and then 54:17 one she had a one shot the introduction 54:20 um and went through that introduction 54:21 and thought it was pretty good thought 54:24 it thought thought it was sort of up to 54:26 snuff. So, I think next we would um 54:30 mindful of where we are with time, 54:33 >> what where would we get to next to then 54:35 get to a full piece and then that that 54:38 Jack would then be taking over to edit 54:41 in his own completely AI native way. 54:44 >> Yeah. So, as I'm writing through section 54:46 by section and having it save things, 54:49 usually I will be working um I will have 54:53 the Google doc open in the inapp 54:55 browser. Um, I didn't do that here just 54:58 because I wanted to keep the screen big 55:00 and have people be able to follow what's 55:02 going on. Um, but as I'm drafting and be 55:05 and getting content that I'm happy with 55:07 or almost happy with, I'm porting it 55:09 over into the Google doc. I'm getting in 55:11 there and tinkering with what I want to 55:13 tinker with manually. Um, and then when 55:15 the piece gets fully drafted, um, there 55:18 are some additional checks that I run. I 55:20 will do sorcin again very often to make 55:22 sure things aren't dragging in places. 55:25 Um I will I have some I have some there 55:27 are some there are official steps in 55:29 that in the pipeline. So, we've got um 55:32 brainstorm, which is the inner or and 55:34 the interview, outline, draft, and then 55:36 we there's a line edit stage and a and a 55:39 or there's a developmental edit stage, 55:41 which I off I sometimes do, and 55:43 sometimes I just swap in the specific 55:45 reviewers I want. But the developmental 55:47 stage looks at the argument and then and 55:50 make sure it's continuous and logical 55:51 and supported. And then the line edit 55:53 goes in and fixes some sentences. Jack 55:56 actually has a lovely skill uh for line 55:59 editing called Titan Draft, which I use 56:02 quite a bit um because um it it does a 56:06 really good job. It it has a really good 56:08 sense that Jack has baked into it with 56:10 his editorial intuition and experience. 56:12 Um it kind of knows the kinds of things 56:14 that we would want to cut. And then I 56:17 make sure all the links and if I'm 56:18 behaving myself, the screenshots get in. 56:21 Um and that's the point at which I would 56:23 pass it to Jack, which I will do now. 56:28 Um, excellent. Uh, so yeah, I'll I'll 56:31 share my screen um here. 56:34 Let me see. 56:48 All right. So, yeah, you should be 56:50 seeing my screen here. Um, so what I 56:52 have here is basically the um the you 56:55 know the the pre-baked version of of 56:57 this um that Katie created. Um and what 57:01 so what I would typically do from here 57:03 when uh you know when a draft from Katie 57:06 or another writer whether it's like our 57:09 other staff writer Laura or someone 57:10 outside of every um an outside 57:13 contributor what I would typically do is 57:16 um the first time through I just want to 57:19 read it from top to bottom. um and just 57:22 kind of have these like bigger picture 57:25 uh thoughts and and comments. Um for the 57:28 sake of you know time and and the stream 57:31 um I'm basically going to kind of like 57:34 combine that like pretend I'm going to 57:36 pretend that like I did that and you 57:37 know the the big picture it looks looks 57:39 okay and and is fundamentally solid. Um 57:42 and then I'm going to go in and I'm 57:43 going to do um my line edit. Um, so even 57:47 when I do my line edit, um, I I think 57:50 you'll find that like I my setup is like 57:52 very very different from Katie's. Um, 57:55 I'm uh I I think like I'm much more of 57:57 like an ultra light hiker in that in the 57:59 sense in that like I I just like try to 58:02 keep everything minimal, but the first 58:03 time I go through what I'm actually 58:06 doing is I want to not actually be 58:11 making like suggestions or making edits 58:14 on the page. What I'm going through and 58:17 doing is basically like reacting to 58:19 things and putting in comments about 58:21 what needs to change without necessarily 58:23 prescribing solutions. Um, so you know, 58:27 so just kind of like going from the top. 58:28 Um, I usually kind of like save the 58:30 title and the, you know, the subtitle 58:32 for for last once I'm like so immersed 58:36 have have been so immersed in the piece 58:37 that I I really understand it and really 58:39 understand, you know, what we might 58:41 title it. So I'll just start from here. 58:44 You know, it is a well-known fact of the 58:45 universe that the Germans have a word 58:47 for everything. Kumers, literally grief 58:50 bacon, is the weight you gain from 58:52 eating your feelings. Backfish, 58:55 I don't know if I pronounce it, is a 58:57 face that's begging to be slapped. Um I 59:00 I this is Yeah, this is hilarious. So, 59:04 you know, I'm just going to comment a 59:06 smiley face. Um, and uh, uh, 59:12 you'll find that like it's like as I'm 59:14 going through because I'm working with 59:16 Katie here, like she knows that in some 59:19 cases I'll be leaving notes for her and 59:21 in other cases I'll be leaving notes for 59:23 either her AI or my AI. Um, so I tend to 59:28 be a little more like brusque with my 59:30 notes and and and for someone who's like 59:32 contributing from the outside who might 59:34 not actually be using an AI agent to 59:36 help them write, I'm I'm tend to be a 59:38 little more gentle. 59:38 >> Basically, you're not worried about 59:40 offending the AI. 59:41 >> Yeah, I'm not worried about offending 59:43 the AI. I'm not worried about offending 59:45 Katie because she knows that sometimes 59:46 like my comments are more meant for like 59:48 Yep. 59:49 >> an AI. Um, so okay. So recently this gem 59:52 came across my ex feed courtesy of Emily 59:54 Cam Campbell, director of AI model 59:57 design at Figma. Um so like sometimes 1:00:00 I'll I'll pick up like um you know 1:00:02 little style things like we usually like 1:00:04 bold um you know bold names and I I'll 1:00:08 kind of like make little changes like 1:00:09 that there. Um so uh Schlimmerbang 1:00:15 literally worsen bettering. It refers to 1:00:17 that most human of all impulses that our 1:00:19 desired to make something better only to 1:00:21 wind up making it worse. Um, can 1:00:23 definitely relate to that used in a 1:00:25 sentence. Uh, not going to try to 1:00:26 pronounce that. Uh, I have worsen 1:00:29 bettered my context. Um, I love this 1:00:32 like um like like great 1:00:36 transition into you know AI topic. Um 1:00:41 and so uh so okay next here like reader 1:00:44 that's precisely what I have done to the 1:00:46 folders on my desktop. So I'm curious 1:00:48 about this because like I can recall a 1:00:52 few instances where Katie does kind of 1:00:56 like direct address saying reader like 1:00:58 this um in working overtime pieces. So 1:01:02 I'm not entirely sure. Maybe that's more 1:01:04 common than than I think, but it's like 1:01:06 it's like, you know, do we typically do 1:01:10 this kind of direct address in working 1:01:15 overtime pieces? 1:01:19 Um, so I I'm I'm basically just like 1:01:21 trying to articulate, you know, what I'm 1:01:24 feeling, what what my reactions are as 1:01:26 I'm going through. um reader. That's 1:01:28 precisely what I've done to the folders 1:01:30 on my desktop, which is inconvenient 1:01:31 because those folders are the main way I 1:01:33 stay sane out here on the decidedly 1:01:36 unpaced AI frontier. Um, 1:01:41 this this is like this is cute, but 1:01:44 maybe needs some more context or at 1:01:49 least a link to uh a post about pacing 1:01:55 the frontier because I feel like yeah, 1:01:57 if if you're not immersed in AI then and 1:02:00 you're coming to this and you're not 1:02:02 aware of what's been in the news, then 1:02:04 then it could, you know, it it it might 1:02:06 not land. Um, so the next one, okay, how 1:02:09 badly did I worse and better them? Badly 1:02:12 enough, that one essay ate roughly 35 1:02:15 hours of my life and produced 91 1:02:17 complete drafts. Yikes. Um, and I 1:02:20 couldn't get any of them right. Um, so I 1:02:24 I see a couple things here like um 1:02:28 like any of them right seems a bit 1:02:34 vague. 1:02:35 Uh, what does right mean in this 1:02:40 context? Can you be more specific? I 1:02:45 probably don't need that last bit. Um, 1:02:46 the other thing I noticed is like, uh, 1:02:49 how badly did I worse and better them? 1:02:51 And there's only a badly here. And I 1:02:52 think it might be neat to actually like 1:02:55 follow the structure of this like 1:02:57 phrase. And so, um, you know, like like 1:03:01 maybe I'll highlight this and say, you 1:03:03 know, I wonder if it 1:03:07 works to like have lines that follow the 1:03:13 worse and better structure. So 1:03:18 like worse enough or like 1:03:22 badly enough 1:03:25 that TK 1:03:28 um well enough 1:03:31 that TK 1:03:33 >> love a TK. 1:03:34 >> Yeah. Um and and for for those who who 1:03:36 aren't familiar with the terms, a TK is 1:03:38 sort of the the kind of like publishing 1:03:41 industry like a placeholder for like you 1:03:44 know something goes here. Um, and the 1:03:45 reason that it's those two letters is 1:03:47 because they don't typically appear in 1:03:50 most in words um, like next to each 1:03:52 other. So, it's it's easier to to kind 1:03:55 of like spot and and and you won't 1:03:57 actually like, you know, see a TK in in 1:04:00 in a real phrase. Um, okay. So, I did it 1:04:04 to myself. I've been handling handing AI 1:04:07 bigger and bigger pieces of my work 1:04:08 because I trusted my context to carry it 1:04:10 through. Um, this sounds like AI. And 1:04:15 the reason it sounds like AI to me is I 1:04:17 feel like AI really loves like overuses 1:04:20 like handing stuff to to other stuff. 1:04:23 Um, and then like something about 1:04:25 >> such a good observation. It's always 1:04:27 handing things over, handing things off. 1:04:29 >> Yeah. Um, so so but but it's like it's 1:04:32 not that alone. It's the handing in 1:04:34 combination with I trusted my context to 1:04:38 carry it through. like something about 1:04:40 those in the same sentence like just 1:04:43 strikes me as like AI sounding and so I 1:04:46 just like mark it like this. Um um and 1:04:49 then uh but I built much of that context 1:04:51 during a hypomomanic stretch. Um 1:04:57 uh so this sticks out because we um we 1:05:02 should give like a little context about 1:05:05 you know like like what is a hypomomanic 1:05:07 stretch? like what how like how does it 1:05:11 pertain to Katie and and Katie's 1:05:13 writing? So, um we should quickly define 1:05:18 or give context here about 1:05:23 Yeah, I think 1:05:24 >> and this is the kind of thing that like 1:05:25 if I had not just oneshotted this from 1:05:28 the previous interview I did, this 1:05:29 probably wouldn't have made it in or if 1:05:30 it had it would have been a little bit 1:05:32 more grounded. But I think like if I did 1:05:35 this piece for real, I would keep it 1:05:36 more broadly relatable about like 1:05:39 hectic, crazy change, um things moving 1:05:43 fast, and we don't need to get into my 1:05:45 particular pathology. 1:05:48 >> Yeah. Yeah. Yeah. For sure. For sure. 1:05:49 Um, and uh, yeah, I I I think, you know, 1:05:54 I think it's like 1:05:56 it's it's so fascinating, Katie, to 1:05:58 actually like see your process live 1:05:59 because because I feel like I've heard 1:06:01 and seen like bits and pieces of it and 1:06:03 and for you to actually like show, you 1:06:05 know, kind of like how you got to this 1:06:07 um this state or or like how you got to 1:06:09 something similar to this um was really 1:06:12 like uh enlightening for me, I think. Um 1:06:16 and so just like very quickly continuing 1:06:19 um uh uh but I built much of that 1:06:22 context during a hypomomanic stretch and 1:06:24 it absorbed my state of mind. So a 1:06:27 comment that I often leave is can we say 1:06:30 this more directly? 1:06:33 Um 1:06:36 uh every experiment got saved, every 1:06:38 correction got promoted to a rule. The 1:06:41 system I was trusting was in effect a 1:06:43 transcript of me at my least steady. 1:06:47 H 1:06:49 system effect like there's something 1:06:52 about this line that's maybe like maybe 1:06:55 a little too lyrical. 1:06:59 Um I don't know. Um 1:07:01 >> I think it leans too much on knowledge 1:07:03 that the reader doesn't have yet. Um 1:07:06 >> yeah, 1:07:06 >> like what like 1:07:07 >> promoted to like correction devoted to a 1:07:10 rule like what does that mean? 1:07:13 >> Yeah. Um here uh yeah like 1:07:21 right this is about comp compound 1:07:25 writing but we haven't set it up yet. 1:07:30 Um, 1:07:32 and then so this is a story of what 1:07:34 broke, how I dug my way out, and what I 1:07:36 learned about shaping your context while 1:07:38 you depend on it. Um, uh, avoid AI over 1:07:43 use of shape. 1:07:47 >> Also broke. AI loves to write about 1:07:51 >> story broke. Um, and I think I think 1:07:54 it's also like this construction like 1:07:56 the story of what broke, how I dug my 1:07:57 way out, and what I learned. Sounds like 1:07:59 AI. 1:08:00 >> It's very like that. 1:08:02 >> I'm very mad that AI took away my rules 1:08:05 of three. Like [laughter] 1:08:07 it's even more devastating than the 1:08:09 dash. 1:08:09 >> You you naturally did that yourself is 1:08:12 what you're saying. 1:08:13 >> Yes. This is a this is a structure I 1:08:15 very often follow and and like it is a 1:08:16 little content marketingy. Um like I 1:08:19 think like my my background being in 1:08:21 content marketing, not journalism. you 1:08:23 see the like I'm gonna make it super 1:08:26 clear for you like what the I'm gonna 1:08:27 map the structure of the of the piece 1:08:29 into the promise and we want to be more 1:08:32 artful than that. 1:08:34 >> Yeah. Um 1:08:36 Yeah. So, so kind of like you know this 1:08:38 intro is an example of how I would kind 1:08:40 of like go through and do my line edits. 1:08:43 Um I'm just going to say let's see. I'm 1:08:46 just to kind of like speed things up. 1:08:48 There's a passage here that 1:08:51 I I think like spotted earlier. 1:08:54 Um 1:08:57 so like for instance like this, you 1:08:59 know, just just for the sake of 1:09:01 demonstration. Um um let's tighten this. 1:09:08 Um and then or actually actually I I'll 1:09:11 save that I'll save that for for another 1:09:13 thing. Um, there's another one here that 1:09:15 like uh 1:09:21 that I spotted earlier. Let me see if I 1:09:24 can find it. 1:09:25 >> We have a comment to Kashik is asking, 1:09:27 "Do you ever flag something as quote 1:09:29 sounds like AI but decide to keep?" 1:09:31 >> For sure. For sure. Yeah. Um I think you 1:09:36 know I think like given the right 1:09:39 context um or like sometimes you know 1:09:43 sometimes we have we have another like a 1:09:46 an AI kind of uh clone of of Kate who 1:09:50 will kind of do a top edit and there 1:09:51 there'll be some things that like you 1:09:53 know that bot will flag that I'm like 1:09:55 actually you know I don't mind you know 1:09:57 I don't mind the use of the word shape 1:09:59 here. Um, and so, so there's definitely 1:10:01 things that like I end up, you know, uh, 1:10:04 wanting to keep. Um, 1:10:07 and so, yeah. So, 1:10:10 for sure. Um, let me let me see if there 1:10:14 was a 1:10:16 Let me see. 1:10:18 Um, 1:10:20 oh, I'm actually going to here. I'm 1:10:21 going to delete this one and I'm going 1:10:23 to say, 1:10:25 um, sounds jargonish. 1:10:29 Um, okay. So, so basically like it's 1:10:32 like if if if we do that then and like I 1:10:36 you know I'll go through the entire 1:10:38 document and kind of leave my comments 1:10:40 all the way through and then from there 1:10:42 depending on a few things like one is 1:10:45 like depending on how much time we have 1:10:48 um in some cases like if you know if we 1:10:50 have plenty of time to edit it I'll just 1:10:52 kind of like let Katie address these 1:10:56 comments without necessarily like 1:10:58 providing suggestions of my 1:11:00 Um if unless like unless you know 1:11:02 there's something that I feel like isn't 1:11:04 clear unless I also give an example. Um 1:11:07 sometimes when we are like on deadline 1:11:10 crunch um I'll sort of basically like 1:11:13 try to leave the ones that I think need 1:11:17 Katie's specific personal experience um 1:11:20 to fill in um in order you know for her. 1:11:24 And then like the other ones, I'll try 1:11:26 to like provide suggestions and to to 1:11:29 help me create those suggestions. 1:11:31 Sometimes that's when I will use um my 1:11:34 AI agent. And so what I found that like 1:11:36 I really really don't like um editing 1:11:41 documents in the cloud code or codeex 1:11:45 browser um just because like it's like I 1:11:47 I I feel like I want separate apps for 1:11:51 different kind types of activities. Um, 1:11:53 and so, you know, more like codelike 1:11:55 things that happens in the the 1:11:58 orchestration apps, whereas like I like 1:12:00 to do my editing just like in this, you 1:12:02 know, I'm using the DIA browser here. 1:12:04 Um, but just straight in the browser. 1:12:06 And um, recently um, in the past couple 1:12:08 of months, I've really been using the 1:12:10 uh, Chat GPT 1:12:12 um, browser Chrome browser extension. 1:12:15 And so what this does is if you have 1:12:18 codeex or the chatgpt app on your 1:12:20 computer, it uses that account and pulls 1:12:24 in everything that that has. So all your 1:12:26 connections, all your kind of like you 1:12:28 know like like uh chat transcripts um 1:12:32 and and and things like that. Um and so 1:12:35 uh so what I'll do here is basically say 1:12:38 like okay you know like review the 1:12:41 comments in uh in this document 1:12:47 and let me know 1:12:50 your suggestions 1:12:52 here. Um so you'll see I'm using uh soul 1:12:56 at medium. Um I haven't really sort of 1:13:00 uh played around with opus 55 enough. um 1:13:04 uh on editing tests that I like you know 1:13:06 really trust it yet. Um but ever since 1:13:10 kind of like you know the the the sole 1:13:12 models and even GPT56 I've just found 1:13:15 that like 56 on medium does a great job 1:13:19 of editing. Um and so so lately I've 1:13:21 been more using so editing. So um what 1:13:24 it does is this whatever tab you have 1:13:26 open um it knows what you're looking at 1:13:29 in the open tab. So if you open a 1:13:31 different tab then like this chat 1:13:34 context gets uh it that that new tab 1:13:37 also has its own chat that is like 1:13:38 unrelated to the to this one. Um so it 1:13:42 says okay I've read all 14 comments um 1:13:44 keep the German uh word opening line 1:13:47 transition those they're positive. So, 1:13:49 this is like it reacting to my smiley 1:13:52 face here, you know, and sometimes it'll 1:13:54 like, you know, uh the agent will think 1:13:56 I'm like, you know, I'm like addressing 1:13:59 it when that's mostly meant for Katie, 1:14:02 but it, you know, no harm, no foul here. 1:14:04 Um, okay. So, uh, in place of the 1:14:07 paragraph beginning, I did um it to 1:14:10 myself. 1:14:12 Um, it says, "Okay, I'd been giving AI 1:14:15 more of my work 1:14:17 because the instructional examples in my 1:14:19 folders had served me well, but I built 1:14:21 much of that material during a 1:14:23 hypomomanic stretch. So, what it's doing 1:14:25 here is it's it's seeing all these 1:14:27 comments and basically like trying to 1:14:29 address them together. Um, and so what I 1:14:33 would be doing is I would be like 1:14:34 reading this and say, okay, you know, 1:14:36 it's like, do I want that? Do I actually 1:14:37 want it to tackle these individual 1:14:39 comments? Um, and if I do want it to 1:14:42 tackle individual comments, one neat 1:14:45 thing about the um, extension is I can 1:14:48 actually just say like highlight 1:14:49 something and when I highlight 1:14:51 something, the selection gets passed in 1:14:54 as context automatically. So uh along 1:14:58 with the comet um it so so it knows you 1:15:01 know it it has access to um the Google 1:15:05 workspace connector and so it's able to 1:15:08 also figure out like what comment is um 1:15:10 attached to uh whatever is highlighting. 1:15:13 So maybe I just want to like address 1:15:15 this this uh highlight for now. So um um 1:15:18 look at the comment here or address 1:15:26 Um, so, so now it's basically reading 1:15:29 the selected passage and its comment and 1:15:32 suggesting a revision here. So, 1:15:36 uh, instead of I'd been handing AI 1:15:38 bigger and bigger pieces of my work 1:15:40 because I trusted my context to carry it 1:15:42 through, it's saying I'd been giving AI 1:15:45 more of the writing because the 1:15:47 instructions and examples in my folders 1:15:49 had worked so well before. This is like 1:15:52 close, but it's like not. It's It's like 1:15:55 I feel like it's better in some ways, 1:15:56 but like worse in other ways. So, I 1:15:58 might be like, "Give me a few more 1:16:01 variations. 1:16:05 >> Big fan of using AI for options." 1:16:08 >> Yes, for sure. 1:16:11 >> I should just say we we do also have 1:16:13 someone in the in the chat who says, "No 1:16:15 love for proof writing. We uh it's 1:16:17 referring to our document uh an AI 1:16:19 document uh tool we have called proof 1:16:22 that we all use for a lot of internal 1:16:24 docs um a lot of internal memos and 1:16:27 things like that but we uh we do when it 1:16:30 comes to writing and editing things for 1:16:32 publication we are wedded to Google Docs 1:16:35 >> on we're wedded [snorts] to Google Docs 1:16:38 um in part because it's like familiar 1:16:40 and also I think like um proof currently 1:16:43 doesn't handle like undo very And so, 1:16:47 >> um, like there are sometimes where I'll 1:16:49 leave a comment and then I'll change my 1:16:51 mind and I'll have to like undo it. 1:16:53 >> And so, as an editor, I use I rely on 1:16:55 that like so much that Yeah. that it's 1:16:57 sort of just, you know, it's like Google 1:17:00 Docs is like second nature to me. 1:17:02 >> Yeah. Editing the edit. 1:17:04 >> Yeah. Um, and so, yeah. So, so it kind 1:17:08 of, you know, it gives me more 1:17:09 suggestions. Um just to kind of like 1:17:11 jump back a little bit like like the 1:17:13 other comments 1:17:16 um you know replace reader that's 1:17:18 precisely what I have done um to that's 1:17:22 what I did to to the folders on my 1:17:24 desktop um it avoids so so see this is 1:17:28 an interesting thing because I feel like 1:17:30 like before like 56 wouldn't have done 1:17:34 this where 56 would have been like let 1:17:36 me let me look at the over working 1:17:38 overtime pieces because I have the every 1:17:42 MCP. It's like I know the site. Let me 1:17:44 read the site, figure out and and so 1:17:46 this is a case where like you know maybe 1:17:50 with six soul um I actually need to like 1:17:53 crank up the effort level because it 1:17:55 seems like it's being a little like lazy 1:17:57 here. Um 1:17:59 >> like it should have told you that 1:18:00 already. It it should know what was 1:18:02 typical of working overtime. 1:18:04 >> Yeah, it should have told me that 1:18:05 already. So, so what what I would do, 1:18:08 you know, what as I'm going through, 1:18:10 it's basically like I'm reviewing each 1:18:12 of these suggestions one by one and 1:18:14 deciding, you know, whether or not like 1:18:17 like it's something that I still want to 1:18:19 implement or some still want to suggest, 1:18:21 whether or not like I can come up with 1:18:24 something better or whether or not I 1:18:26 want to like take the AI suggestion. 1:18:28 [snorts] Um, basically like kind of like 1:18:30 doing that through, you know, the entire 1:18:33 document and probably like right now 1:18:35 like like editing a piece like this from 1:18:37 top to bottom um would probably I would 1:18:40 guess it would take me like three to 1:18:41 four hours. Um like kind of like you 1:18:44 know following this this kind of um 1:18:47 pattern. Um I'll show you a couple other 1:18:50 things that I have. So, um Katie 1:18:54 mentioned um a skill that I've set up to 1:18:58 um tighten uh drafts. So, um what that 1:19:03 skill does and I'll show you the the 1:19:04 skill text, but I can just kind of like 1:19:06 demo this. Um so, let's like let's 1:19:10 highlight this. And so, I'll just like 1:19:14 tighten draft. 1:19:17 And that again, you know, it kind of 1:19:18 like pulls in the selection. 1:19:22 Um, 1:19:29 and basically like the the this the 1:19:32 skill is kind of it's like applying the 1:19:34 um the Stephen King rule which is that 1:19:36 like uh you know from Stephen King's 1:19:38 book on writing which is saying that 1:19:40 like a second draft should be a first 1:19:42 draft minus 10%. Um it's like partly 1:19:45 doing that and then like just like 1:19:47 partly trying to you know there there's 1:19:49 also kind of like suggestions for ways 1:19:51 to tighten that I have in the skill. Um 1:19:56 so 1:19:59 let's see. 1:20:02 Yeah. So keep the three instructions uh 1:20:05 instruction excerpts and the great 1:20:06 Gatsby minus the party's line. Um and 1:20:09 then like you know change the records 1:20:11 piled up. by the time I stop my desktop 1:20:13 held 91 separate instruction files just 1:20:15 to by the time I stop my desktop held 1:20:18 you know to cut that first part get to 1:20:20 the failure sooner like what I didn't 1:20:22 see was that my system couldn't tell a 1:20:24 record from a rule um you can just state 1:20:27 that more directly my system couldn't 1:20:29 tell a record from a rule so it kind of 1:20:31 does that and then it tells you like you 1:20:34 know it's like the the revised version 1:20:36 is like you know the original minus uh 1:20:38 yeah uh cut from 172 2 words to 145 1:20:42 words and then like it's trying to shoot 1:20:44 for that like 10ish% target. Um so the 1:20:49 other one that I have um set up um is 1:20:53 one that is um called the jargonify. 1:20:58 Um 1:21:00 so let's see the example I had here 1:21:05 um is that like every experiment every c 1:21:10 so so let's just do it to this one I 1:21:11 I'll have like 1:21:14 but basically what this does it it's 1:21:17 there's a little bit of overlap with 1:21:18 Titan but it looks for technical 1:21:20 language like maybe like here it would 1:21:22 pick up you know corrections getting 1:21:24 promoted to rules um and it tries to 1:21:27 state it in plain English. Um, and then, 1:21:30 you know, I saved our experiment and the 1:21:32 system treated each correction as a rule 1:21:34 for future drafts, which is like a lot 1:21:36 more legible to like reader. Yeah. 1:21:39 >> Um, so I'll I'll kind of show you these 1:21:42 um these skills here. Let me 1:21:45 um share. 1:21:52 Um, so basically here's the Titan draft 1:21:55 skill. And so it says, "Aim for a, you 1:21:58 know, 10 to 15% word count reduction 1:22:00 when the draft supports it. Stop sooner 1:22:02 if further cuts would weaken meaning, 1:22:04 voice, rhythm, or necessary context." 1:22:07 Um, this is saying it's like don't make 1:22:10 direct edits to the document. Just like, 1:22:12 you know, uh, tell me like the suggested 1:22:15 changes in chat. Um, and there's a kind 1:22:18 of a workflow. It's like read the full 1:22:19 draft before editing because you want to 1:22:21 know where the passage sits in the 1:22:24 context. cuts structural bloat before 1:22:26 tightening sentences where you know 1:22:28 sometimes like a paragraph might not 1:22:30 even need to be there. So I want it to 1:22:32 do those like larger levels. Um and then 1:22:34 it kind of like applies the rules below 1:22:36 and then gives me like a readout of um 1:22:39 so like first like you know it'll flag 1:22:42 paragraphs that are worth cutting 1:22:44 altogether. Um and then like you know 1:22:48 like what characterizes a paragraph 1:22:50 that's worth cutting. Um and then here 1:22:53 are editing rules like pre prefer 1:22:55 concrete detail over you know abstract 1:22:57 labels. Um remove like throat clearing 1:23:01 sentences um bridge sentences like it's 1:23:04 worth noting or this means or in other 1:23:06 words. Um expose the subject and verb 1:23:08 you know get rid of the there is and 1:23:10 there are and it is u make the verbs 1:23:12 more active like cut the adverbs. Um 1:23:16 prefer like present or past tense 1:23:18 instead of jirens or you know 1:23:20 participles. 1:23:22 Um, and then like you know abstract 1:23:24 closing sentences that are like like 1:23:27 from X to Y when um you know when you 1:23:31 can end on a concrete claim. Um, so 1:23:35 that's kind of the uh you know the the 1:23:37 the Titan draft and then the the 1:23:39 Jargonify 1:23:41 um also has its own kind of like 1:23:43 workflow where first I have it like 1:23:46 remove jargon from the sentence um and 1:23:49 then basically like like then 1:23:53 I noticed that sometimes when I would 1:23:55 ask my agent to like remove the jargon, 1:23:58 it would introduce a bunch of like other 1:24:02 phrases that would then need to be 1:24:04 tightened. So then it has a tightening 1:24:06 pass. And then like what I found doing 1:24:08 that was that sometimes it would end up 1:24:10 like basically like repeating an idea 1:24:13 that's elsewhere in the sentence or that 1:24:15 like it would change that particular 1:24:16 sentence, but then the flow and 1:24:19 transitions in and out of that sentence 1:24:21 would not be that great. So then I have 1:24:23 it kind of like reread the whole whole 1:24:25 text for flow, which is basically like 1:24:28 what I would do is like if I'm changing 1:24:29 a sentence after I make that change, I'm 1:24:32 going back and like rereading the 1:24:34 sentence in context to make sure that it 1:24:36 like it it continues to flow. Um and 1:24:39 then so you know this is a very short 1:24:40 skill like then like you know the output 1:24:44 format for each suggested edit. 1:24:47 Um, so, so yeah. So those are basically 1:24:51 like the the the the two skills that I 1:24:55 use. The the only other one that I use 1:24:57 personally is one that is like a 1:25:00 proofreader. That's basically when it 1:25:02 comes to production. Um, sometimes like 1:25:05 we're copying and pasting the Google 1:25:07 document into our CMS and it introduces 1:25:09 these like line breaks um, for whatever 1:25:12 reason and so the proof reader will like 1:25:15 catch that. Um, and then I think like 1:25:17 before we end, I can show you one more 1:25:20 thing. Um, which is basically 1:25:24 um what I would do before handing this 1:25:28 to Kate for her top edit, which is like 1:25:30 kind of the final edit before it goes uh 1:25:33 live on the site. Like Kate, you know, 1:25:35 edits is like the last person to 1:25:37 basically like touch every single piece. 1:25:39 Um, and so this is 1:25:44 um, 1:25:46 so this is our um, this is the the 1:25:50 thread that um, uh, Katie initially like 1:25:53 shared in our company Slack um, for like 1:25:56 posting the the draft. Um, and so like 1:25:58 like let's pretend I've done my pass, 1:26:00 you know, Katie's done her edits address 1:26:02 and we've we've basically like it it's 1:26:05 like ready for Kate. So what we do 1:26:08 before sending it to Kate um you know if 1:26:10 we have time sometimes we we don't 1:26:12 always have time um but um I will tag 1:26:15 every which is our every agent um uh and 1:26:19 I'll tell every to run uh 1:26:24 to run Katebench on this document and 1:26:27 what Kate Bunch is is sort of um 1:26:30 basically like Kate's editorial tastes 1:26:33 uh her like copy editing tastes um that 1:26:36 have been um you know that we've we've 1:26:40 created kind of the skill and this like 1:26:42 benchmark to try to like replicate as 1:26:44 much of that as we can with the thinking 1:26:47 that like you know this way by the time 1:26:49 it gets to Kate like some of the obvious 1:26:52 things that need changing well you know 1:26:54 I'll have like kind of like picked up on 1:26:56 those that I might have missed in my 1:26:57 editing pass so that when it finally 1:26:59 gets to her you know she can focus on 1:27:01 like the big picture like questions like 1:27:04 you know does this like meet our 1:27:05 editorial standards like are there any 1:27:07 like big things that need to move rather 1:27:09 than like kind of like these like 1:27:10 smaller like um you know nitpicky things 1:27:13 that are very much like rulebased like 1:27:15 you know we don't put spaces around our 1:27:18 m dashes and things like that. do not. 1:27:20 >> Um, and so, uh, I think like we're 1:27:23 getting close to the end of the time, 1:27:25 but but what this does is it it it'll 1:27:29 take all what it finds here. And, um, 1:27:32 Katebench actually has um, uh, access. 1:27:35 There's a we have a um, a Google Doc 1:27:39 employee called Kate assistant that um, 1:27:42 Katebench drives. And then Kate 1:27:43 assistant will actually go into the 1:27:45 document and as kind of as track changes 1:27:49 um basically like put all those 1:27:51 suggestions as track changes throughout 1:27:53 the document. Um and then what I'll do 1:27:56 is I'll go in and I'll kind of like mark 1:27:58 the ones that are either clear rejects 1:28:00 or clear accepts. And the ones I'm not 1:28:03 sure of I'll I'll leave for Kate. And 1:28:05 then I'll like I'll you know I'll let 1:28:07 Kate know that um that this draft is uh 1:28:10 is then um ready for her to take a look. 1:28:18 >> You're muted. 1:28:19 >> Just a Okay, I'm unmuted. Thank you. Um 1:28:22 just a couple of notes. One, you saw 1:28:24 that Jack tagged every meaning the every 1:28:26 agent. That is um an agent, a company 1:28:30 agent that we have in in beta that we're 1:28:32 going to be releasing in a couple of 1:28:33 weeks. So all of you can uh can can can 1:28:37 get it in your own Slack as well. Um so 1:28:40 look out for that. Um and also just a 1:28:43 just a note on Kate Bench which I think 1:28:44 we've talked about at various points. It 1:28:46 probably takes about um it can take up 1:28:49 to about 10 minutes to run depending on 1:28:51 the length of the piece. So it is 1:28:53 something that you're you know isn't you 1:28:55 know I think we're all used to sort of 1:28:56 AI working immediately. Um this is 1:28:59 something that just takes a little bit 1:29:01 longer. So it's the kind of thing we set 1:29:02 off on a run and go do something else. 1:29:06 Um and so uh and so then come back to it 1:29:09 and we see we see the document. Um so um 1:29:13 I thank you all. Thank you first of all 1:29:15 Jack and Katie for taking us through 1:29:16 your process. Uh we know that that's uh 1:29:20 narrating that is a lot of work and 1:29:22 really appreciate it. Thank you everyone 1:29:24 for joining. As a reminder we are every 1:29:26 every two the only subscription you need 1:29:28 to stay um at the edge of AI. We have 1:29:31 one last question I should say is how 1:29:32 happy are you with the results? Uh I 1:29:35 don't know if either of you wants to 1:29:36 chime in on that knowing that this was a 1:29:38 test case, not a not necessarily a real 1:29:40 a real one. 1:29:42 >> Well, you you will get to read a version 1:29:44 of this article soon hopefully, right? 1:29:46 Like that that's the plan, right? 1:29:48 >> I'm not all this work. [laughter] 1:29:51 Um, no. I mean, I was really happy with 1:29:54 the the intro that we wrote was so like 1:29:56 on live was so much better than the one 1:29:57 that I just yeated out, you know, in 1:30:00 order to get Jack something to to edit. 1:30:02 But, um, this is this is how it works 1:30:05 when it's working well. So, yay, I 1:30:07 figured out my context. 1:30:10 >> Great. Well, thank you everyone for 1:30:12 joining. Uh, and we will see you again 1:30:14 soon.