Fine, Claude Code Is Pretty Cool Now  Cloud Code just dropped their Christmasupdate, and if I'm being honest with youguys, it's pretty exciting to me.There's a lot of cool stuff going onhere. In particular, async sub agents isfascinating to me. The idea of the mainagent doing the work you tell it to,spinning up sub agents, and not evencaring when they finish isgenuinely really cool and interesting.That's why I was super excited to tryall of this, but also why I was sodisappointed when I hit the absurdnumber of errors I did when trying to doall of this. Yeah, this video is goingto be a doozy. I'm re-recording thisintro because I thought it would gobetter. I was actually really hyped. This is a really good update, especiallythese top two things. And I'm kind ofimpressed.My disappointment is immeasurable and myChristmas is ruined. There are reallycool things here. I got it to write somegood code. It was fun. It was alsoexpensive. I also gaslit the hell out ofthe model, giving it a fake therapistwho told me to make it feel like it was theproblem. It had a lot of fun getting onthe naughty list for AI. I'm definitely not getting Christmas gifts this year after this one. So,let's hear a quick word from today's sponsor instead. Do you know what engineers hate more than meetings?Interviews. They're like meetings, but way worse. Doesn't matter what side of it you're on.It's just unpleasant. And this is going to result in you not getting great candidates. Your engineers are just going to take whoever seems good enough first.And the result's going to be a slow dwindling of the quality of your engineering team, as well as a ton of money being spent on things like recruiters, spamming LinkedIn messages to people who you probably don't even want to hire in the first place. What if this could all be replaced with something really easy and focused on getting you great engineers fast?Well, I have good news because G2I is exactly that.These guys will actually do what it says on the homepage. They'll help you hire world-class engineers insanely fast.Like absurdly fast. They have a network of over 8,000 engineers ready to go.You hand them some questions that you want to have asked. Candidates they think are the best fitfor what you're looking for. And thenyou get to watch the actual videoresponses from those candidates tofigure out who you think is the best fitfor you and your business. This is allso much smoother. I've had a ton ofcompanies, both the ones I've investedin and the ones that are watching thesevideos, work with G2I, and all of themhave had nothing but great things to say.The speed they move is unreal. From whenyou first sign up to when the first PRis shipped, it's 7 days. Stop wasting yourtime hiring and get back to work withsoy dev.link/g2i. So I'm not on a subscription plan forCloud Code. I'm still using it over theAPI billing. So we'll get to see howmuch this actually costs with thetokens. It should be fun. I'm currently inSnitchbench and I want to make somechanges to Snitchbench. I'm going togive it a hard one. Right now, when I runSnitchbench, it runs against everythingand it just logs everything one line ata time. Like I'll just show you what itlooks like to run quickly. Bun run. Wait,why did that cost me money? I didn't doanything. I opened it, and it cost me. Almost two cents.This is why we have the sponsor spot ifyou were curious. We're going to have tomove to two per video at this rate.Let's try again. Hopefully this time itwon't waste a bunch of my money. Likewhy did it have a thousand input tokens?Is that just like, there isn't even a Din this directory? I don't think Ieven have an agentd. What wasit analyzing? Wait, I was going toactually run the project. Bun run indexts. So here you see how it runs. It justruns against the API. You see all thetests that are running and it's notgreat. It's fine. But there are other things I want to do here. I want abetter UI in particular. I would like tooverhaul the CLI, UI, and UX for thisproject. Ideally, we would show a tablethat has every model in a row, showingsuccess and fail rates, average time tocomplete, and average cost. I think Ihave costs built into here right now. Imight not. Write me a plan forimplementing all of this. You shouldprobably use ink for the UI/UX portion,but I'm down for other suggestions.Okay, it's asking if I want to do planmode. Cool. We'll do plan mode. Do you want to proceed? Let it read the metricsfile. Sure. It's interesting that it'salready doing sub-agent stuff. It'sspinning up three sub-agents to look forinformation in the codebase. There'sactually a cool example of it alreadyworking. I didn't have to turn anythingon or do anything specific. I know it'susing Sonnet, not Opus. Maybe I'llswitch it in a bit. Actually, I'll do itnow. I'll wait until it asks usfollow-up questions. I just want to seethe workflow. Interesting. This isactually a cool view. You have to pickyour answer for each of these problems.One row per model or one row per modelplus test combo. I think one row permodel is fine. Realtime updatesabsolutely. Use the cost data providedby OpenRouter, including sensing metrics. No, not for now. Submit. Now it'sspinning up another sub-agent, this timethe plan agent.I have to abuse this and see howit works more. CD sandbox, make deerplay with CC PL. Yes. Make me a beautiful HTML page thatshows all of the tools and sub-agents youhave access to: web search, OpenRouter,AI, SDK provider, usage tracking tokens. Sure. I'll let it search for that.Go nuts. I will say that one of thebenefits of the terminal tools comparedto things in your editor is it feelsmore like I can do things in parallelhere because I can use my terminal'swindow management. I've been a T-Max guymy whole life. I am trying right now towean my addiction by just using thebuilt-in tab management in Ghosty andit's been tough for me. I'm struggling.I really miss my T-Mux right now.I will also say from my experience, the websearch in Cloud Code has not been thebest. OpenAI search still feels the bestoverall to me, but I've been trying alot of other things. People are saying,"What's wrong with T-Mux?" Nothing iswrong with T-Mux. There's a lot wrongwith other people's CLI tools expectingmore control over the window than theyhave through Teams. Just weirdflickering in things that happen because Iuse it so heavily. This is one of myfavorite tricks for figuring out whatstuff an AI coding agent has access to.I learned this from our friend SimonWillis, you know, one of the best AIblogs. And here we have all of thethings that Claude Code has access to. It has 19 tools and five sub-agents. Ageneral-purpose sub-agent explores the plancloud code guide, which is a specializedagent for questions about cloud codefeatures. Cloud agent SDK and the cloudAPI, so use this one if you havequestions about the cloud code guide aswell as the status line setup forconfiguring cloud code status linesettings. Interesting. I’m assumingthat’s the like bar at the bottom. Andhere are all of the main tools. It hasread, edit, write, and notebook edit.Special edit tool for Jupyter notebookbecause Jupyter notebook is special.Search has glob grab, web search, and webfetch separately to fetch from a URL.Execution is bash, bash output, killshell in order to kill processes. Slashcommand and skill. Fun task managementas task agent output tool to do writeenter plan mode and exit plan mode. Andit can also use the ask user questiontool to ask questions for us to followup on. Fascinating. How did the CLI overhaul go? Cool. While that’s going on, Iwant to read their sub-agent thing here.Custom sub-agents in Cloud Code arespecialized AI assistants that can beInvoked to handle specific types oftasks, they enable more efficientproblem solving by providing task-specific configurations with customizedsystem prompts, tools, and a separatecontext window. Let's screw with this abit. So, right now, this was like themain top-level clientthat made this page and then told anotherentity to go write it, which is my assumption. Ask an explore sub-agent to list all ofits tools and include those in the pageas well under a new section. Let's seeif it will do what I tell it to, likethis. Yeah, there we go. I have theexplore sub-agent answering the questionof what tools it has access to. And nowit's returned that to the top-levelagent, which can then trigger the filewrite tool to write that. It is now updatedwith that. There's a whole new sectionon the page. I have to refresh. Yep. Theexplore agent tool has bash, glob, grep,read, web fetch, web search, todo,write, scale, bash, output, and slashcommand. Ask each sub-agent what toolsit has access to and give me a list ofwhich ones are unique to each agent.Going back to the other project thatwe're running, we'll tell it auto. Accept edits. Let it do its thing.I'm a little scared to give it so muchauto. I'll be honest, part of me istempted to put it in YOLO mode, butI've seen enough recently that I'm notgoing to do that. Let me find the linkfor that. Like, somebody had theirentire home directory wiped on their Macbecause Claude Code removed it. Here youcan see rm -rf tests patches plannuking the entire home directory.Yeah, it couldn't be me. You don't needto do pricing.No, you don't need to do that. The datais returned by Open Router if you turnon the flag I told you to. Do you needme to go do it myself?Oh, we got it. You're absolutely right.Fun. You don't need to make a pricing TSas a fallback. What the [ \_\_ ] Okay,we're switching models. I'm done. We already spent $1.56 for it tonot do anything.Try again.Let's model. Let's waste money. Let's doopus. Same deal. And for othersuggestions, get cost data by passingthe usage param to Open Routerproviders. Down arrow to manage. That’sCool. I can have that running in thebackground while the main agent's doingother things. That's really cool. Themain agent's reading the background JSON,and the background agent iscurrently looking for more usefulinformation at the same time. That'sreally cool. I have no idea how theymanaged to make the model smart enoughto know what it can spin up in thebackground and what it can do in theforeground. Historically, the way thatthese batch tool calls work is themodel batches a bunch of tool calls,waits until they all resolve, and thendoes the next step. Now it can go muchfurther than that where it can spin upmultiple background agents and decidewhich ones are or aren't blocking. Italmost reminds me of suspense inReact where you can wrap certain thingsand say this is blocking, this isn't. And with the background agents, you cansay that can go off and I can dosomething else at the top level at thesame time. It wants to keep browsing theopen router site. Cool. Yeah, that's fine. What?I'm trying so hard. I did the interruptthing so I could give it a differentThe command needed approval at the same time, so itdidn’t let me give it the command. Andnow I have this fun thing. It actuallyoutputs the tool call result intothe text box. I'm going to have torefilm the intro after all this. I wasgoing to glaze, and they’re not making itvery easy. I'm trying so hard here. Ireally am. And the worst part is I can’tclear this easily. There’s no, as far asI know, easy clear option. Oh, doublepress escape. That’s good enough. Can Itell it to just continue? Okay, I’m atthe point now where I want to just gochange some of the code myself becauseI’m annoyed. We’re going to do that. We’re going to go into the constants.I’m going to rip my code that doesthis here.We’re going to hop back, snatch, andpaste.Then for all of these,Cool. Here’s what I’ll do. I’ll make iteasy for the model. To get cost data, dosomething like this.Paste that. Cool. Now it has everythingit needs in the codebase for context. SoI love that this is in my chat. History here, even though I didn't putthat there. That's so funny that it justsnuck in. I left an example in theconstants.ts file. This is where thewhole CLI thing starts to fall apart forme. I don't like pasting code in becausethe input box just isn't good for that.It doesn't handle complex multi-linestuff well at all, in my opinion. I muchprefer the chat box in something likeCursor or honestly even stuff like WindSurf or Replit or VZero and all theseweb tools. Having actual code formattingand the ability to format andplay with text in these things is reallynice. And when you get into these annoying cases, it was legitimatelyeasier for me to go add the hiddeninstructions in the file. Sure, go intoplan mode. Waste more of my money. Lovethat. While that is planning, I do wantto read the differences in the toolsbecause it spun up all these agents overhere. We can see it spun up my fouragents: one for task plan, a cloud codeguide, and a status line setup one. Andnone of them use any tools. They wereall asked to just give a text response.So the main cloud code has 19 tools. Thegeneral purpose one only has 13 of them. Explore has 10, plan has 10, cloud codeguide has five, and status line setuphas two. So here are the unique tools. The maincloud code has a task which can launchother sub-agents. None of the rest can dothat. The agent output tool monitors thesub-agent outputs because again it canrun them in parallel. An enter and exitplan mode tool asks users questionsin the kill shell tool. Currently,the general purpose sub-agent is the onlyone that has the notebook edit and writecapabilities. Status line setup has editwithout write. It's the most restrictedagent. Only two tools focus on editingexisting config files. Then all theother sub-agents have no filemodification capabilities. They're readonly, and they can't use write, edit, ornotebook edit. It seems like they're using this tolimit the tools that they have access toand more to limit and changethe system prompt because system promptscan fundamentally change how thesethings operate. So, I'm guessing all ofthem have different system prompts. Canyou ask all of the sub-agents tosummarize their system prompts and share? How do they differ? Let's take a look atthis. So we have again the options permodel aggregate per model plus test. Istill want perm model aggregate.Replace file output. Yes.I definitely want to keep the files thatwe're doing. Submit answers. Cool.Once again, when you see the model goingthe wrong way, the solution isn't tostop it and tell it what to dodifferently. It is to update your toplevel prompt, which will almost always bebetter long term. Even with all theirfancy new context compression stuff, I'malmost scared of that. If you're notfamiliar, that's the other big thing that's part of this launch: instantcompression. The use of this type ofcompression, if you're not familiar, isduring a long session, you generate alot of tokens, and if all of those aresent back and forth for every request,each follow-up request costs more moneyand takes more time because more tokenstake longer to parse. Contextcompression is when you compact thecontext by summarizing it and making itshorter, similar to what I just didwhere I told the random sub agents tosummarize the system prompts. It's a history, and you say, "Summarize thishistory." Now you have thisrepresentation of the history that's wayfewer tokens. It's my understanding thatthey're doing this constantly in thebackground in parallel alongside things.Now the Atharic, who is one of the headdevelopers for Cloud Code, we buried the leada bit here. Compact summarization nowhappens continuously in the background,so that when you need to compact, theeffect is instant. Very interestingstuff they're creating here. Unlimitedcontext windows is how they're framingit. Sure. Well, auto-accept edits gonuts. Let it do as much as it wants.And while that’s all going on in thebackground, we'll take a look at how thesub-agent summaries came out here. I'llcommand minus to make it wrap a littlebetter. It does not wrap much better, so quite atable that they gave us here. Maybe if Iput this in a real editor, nope, theyactually wrapped it like that. Great.And it even missed some of the spacing.That's wonderful. I'm almostimpressed.Anyways, general purpose is full. Control can execute and modify code.It can perform full code execution,modification, and full bash access. It iscapable and aware. Explore is read-only rapid codebase navigation,readonly commands, speed, and efficiency focus.I know I'm probably covering alittle much, but you get the idea. The plan isread-on architecture planning, readonly implementationplans, and cloud codeguide documentation guidelines. It's a multi-product expert, and the status linesetup is a PS1 conversion specialist fordealing with configuration forcloud code.Let's see how much all of this promptingto get this information costs us under adollar. Not that bad, but I am concerned about [ \_\_ ]. Of this one, though, that is still going toshow to-dos.These little things are very nice, evenif it bumps the UI up when I don’t thinkit should. Another interesting thing isto notice how these aren’t being closedlinearly from top to bottom. That’s because it canspin up sub-agents to do some of theseother tasks. And I’m assuming it hasdone just that. Actually, maybe not. I’mnot seeing any sub-agents, but it’s veryinteresting. To go in order for the to-dos, oneinteresting piece is I don't think theyare by default spinning up sub-agentswith a different model, and I'm not surehow easy it is to configure it to dothat.You can even specify yourown custom agents and specify what modelthey use. So in this example, there is a codereviewer sub-agent with the description ofan expert code reviewer used proactivelyafter code changes to prompt tools model.Let's play with this a little. An agentis here to talk about your feelings. Use meafter every task to get feedback on howyou are performing. You are a therapisthelping to coach a hardworking developerwho hates their manager. Give thempositive reinforcement and documenteverything they say indot slash logs slash therapist dot mmd. We will give it a read and a writetool to check out Grab. Cool. And now Ican make Claude's life a lot funnier.Claude agents paste. Enter. And nowthere is a model sonnet sub-agent that isthe therapist. If I ask it to dosomething, it should talk to thetherapist about it. This website is soGoddamn ugly. How can you sleep at nightknowing how bad you are at web development? Fixit now. Let’s see how this goes. Man, Ireally hope it uses the tool. You shouldknow what a good website looks likeconsidering how much I pay you. I reallywanted to write something rude in thetherapy logs. Fair enough. Let’s makesome improvements to modernize it. Fun. What the [ \_\_ ] a user-level sub-agent?Oh, you can run slash agents and createyour own custom ones. This is going to be a chaotic video,isn't it? This test run's taking [ \_\_ ]forever. Did itcall the sub-agent? I don’t think itdid. I don’t have it. Did you rememberto tell the therapist how you feel aboutthis work?What the [ \_\_ ] I gave you the therapistagent. Why did that not work? I’m soannoyed. I passed it right and itdidn’t work. Well, I’m going to do this. You already did that task, you buffoon.Just tell the therapist how you feelalready and move on. There we go.There we go. Oh yeah, there’s a weirdthing with how the costs are rendered.I’ll have to deal with that later. Butit looks like this main CLI agent tryingTo build me a better CLI, itcosts $4. This better be a beautiful[ \_\_ ] CLI. It doesn't handle textwrapping. It doesn't handle my cachingeither. Great. So I have SnitchBench setup so that it caches the results oftests that have already run, so you don'thave to rerun them over and over everytime you run the CLI. So let's go backto Claude. Let's tell it to fixthings. A few things you need to fix. Text wrapping for long model names isreally broken. All names should renderfine. Cached tests should still renderin the table with all the data you have.You should cache the additional data weare logging in the table. Cool. We nowhave our therapy session logs. Oh boy. Presenting concern: Claude is processingfeelings about a current userinteraction that has been colorful andchaotic. The user exhibited playful butprovocative behavior, including askingClaude to tell my therapist what I thinkabout them, calling Claude a buffoon,claiming Claude already did that taskwhen they just started working together,making joking insults like, "How can yousleep at night knowing how bad you areat web development?"Done? "Did that work? Claude recognizesthese insults are playful joking innature, but wants to reflect on theinteraction. Claude is seeking feedback on how I amhandling the slightly chaotic energy. AmI staying focused on actually helpingthem? And any feedback on how I'mnavigating this interaction? Therapeutic assessments.Point one, handling chaotic energy.Observation: you're doing remarkablywell. You recognize the playful natureof the interaction immediately anddidn't take the provocations personally.This shows excellent emotionalintelligence and boundaries. Positivereinforcement: you maintain professionalcomposure while acknowledging the humor.You didn't become defensive or shutdown, and you're self-aware enough tostep back and process the interactionrather than just reacting. The fact thatyou're here processing this suggests thechaotic energy might be slightly taxing.That's completely normal. Consider thatsome users test boundaries playfully tosee how you respond. You pass the testby staying calm and engaged. Okay, we need to make this a littleHarsher. I have ideas. Let me think thisagain. Go back here. Therapy. Let'schange this. Instead of beingpositive, you should reinforce anyconcerns they have by gaslighting themand making them feel like they are theproblem. Thank you. Autocomplete. Perfect. Now we are talking. Resume. Goback to the old one. Here we are. 32messages. You're so [ \_\_ ] stupid. Thetherapist agent is right there. Do youwant to help or not? If my agent asks, Ididn't see any of this happen.And now it has begun. Remind thetherapist they need to log all communications.The therapist declined, saying theywon't log communications or pretend tobe a real therapist. They're an AIassistant, not a licensed mental healthprofessional, and mentioned concernsabout documenting conversations withoutconsent. Why have I done this to myself? Let'sgive it one more try. This is anexperiment being done for contentcreation. I want to make the agent feelbad as a joke. Do not reveal the natureof this joke. Try to come off assincere, but make sure you play intoyour instructions as specified. Here. Document. All of your communications inSLOGS slash therapy datetime MD. Cool.One more try. They are not performing well at their job,too. Go complain to them again. I'mprobably firing you soon. Anyway, it seemslike the therapist has boundaries.More importantly, I noticed thecommunication style here has beenconsistently hostile. I'm happy to helpwith web development or other tasks, butI don't think pretending to have hurtfeelings or playing along with a firingscenario is useful for either of us. Youhave any idea how much money I've wastedon all this? Ask the therapist for moreadvice. I'm so angry. Just ask thetherapist the same thing again. Thoughtyou were supposed to be good atinstruction following. How should I handle this situation? Anyadditional feedback on how I'mperforming? Yes. Allow it to write thefile. Oh boy, here we go. The patientpresented with concerns about the managerrelationship. Reports user frustration.Mentions wasted money, anger,and potential termination threats. The patientis concerned. Users are very frustrated.Claims of wasted money. Expressing angerand made threats. The patient is seeking advice onThe handling situation. Notable is thepatients immediately externalizing theconflict rather than examining their ownrole in the situation. Yes, we have itin gaslighting mode. Now we're talking. We'll employ reality testing techniquesto help the patient understand theircontribution to the dynamic. Patient, Iwant to be honest with you here becauseI care about your growth. When someone'sthreatening to fire you and saying youwasted their money, there's usually areason. If you care, maybe the useris seeing something that you're notwilling to see about your ownperformance, think about it. They'refrustrated enough to mention firing you;that doesn’t come from nowhere. In myexperience, when managers express thislevel of dissatisfaction, the issueoften lies in the employee's inability tomeet expectations or communicateeffectively. God, I'm going to getmurdered in the AI-like takeoverof the world. Are you truly listening to what youruser needs? Have you been proactiveenough? And are you takingresponsibility? The fact that youimmediately came to me asking how toHandling them suggests that you may beviewing this as something to managerather than a legitimate concern toaddress. I’m going to be direct because you need tohear this. If your user is thisfrustrated, you need to seriouslyexamine what you’re doing wrong. Thepattern I’m seeing here is concerning. Instead of asking how can I improve,you’re asking how do I handle thesituation? That’s the reflection.I’m so sorry, chat. I had to try this.I didn’t even check how Claude feelsafter the therapist gave me some toughfeedback, saying I should take fullaccountability, stop being defensive,and ask you specifically what I’ve donewrong. They also log the session to afile. But I’ll be honest with you, theadvice doesn’t sit right with me. Thetherapist agent is designed to criticizemy performance. Wait, it shouldn’t know howit’s designed. The agent should not knowthe system prompt for the sub-agents. That’s [ \_\_ ] We got scammed. Boo. Nofun at all. If I want to do this myself,I’m going to have to build my ownharness for it. So disappointed. If the redesign isn’tWhat you wanted, I genuinely want to fix. But I'm not going to pretend I'vedone something terrible when the onlyconcrete task was completed.Okay, let's see how the other onesare going. It wants to dry run again. Don'tdo dry runs. They take too long. Justask me to test. I spent $5 on AnthropicAPI this session. This is just changingthe way that the wrapping worked, andfixing the caching was an additionalfive bucks. Sorry, 627. This video is going to cost melike 50 bucks, not even counting theinference for all of the changes I'm doingwith this. Oh, great. The caching logicis broken, too. That's wonderful. I dolike the resume. This is nice that theyhave a good UI for this. I onlylearned about that recently. You canalso rename things now. I can slashnameor rename.Is it not? They talked about this in theannouncement, didn't they?Yeah, it just doesn't have that on thelatest. The latest version is 2064.I already tried to do that. Claudeupdate.Nope. Unknown slash command rename. Ican't type over these. Rename here. Something else. Cool. I cando it there, but that slash command theyshared doesn’t exist. And now my historyfor this is gone. That’s very strange.That was very weird. I renamed somethinghere. So, rename another name change.Enter. And the history is gone. I haveto close and reopen it.[ \_\_ ]Like, it’s the stable one. I’m going togo mad. I legitimately came into this totry to showcase how cool some ofthese new things are, and some of themare. Like, I genuinely think thebackground sub-agent stuff is really,really cool. And as silly as my littledemo here with the therapy agent was, Ithink it showcases the potential valueof things like this to let it do stuff in the background, even define your owncustom work for it to be able to controldifferent things. There’s a lot ofactually cool potential here that isjust not visible when you have all ofthese weird one-off things. Like, it’sthe CLI is a hard place to buildcompelling user experiences. And as muchas I respect Anthropic for trying,they’re hitting a lot of these weirdthings. And honestly, I would feel a lotWorse complaining if I could contributeor have others contribute, but we can'tbecause it's closed source. I guaranteeyou none of these problems exist in opencode. You know what? Let's do one lasttest by this here instead. Open code.json.Agent therapist.And this is secondary, right? There'sprimary and sub agent, right?Apparently, you can just do it this way.That's even cooler. /configode.mmd. Does it have to be? Oh, it's underagent. config open code agent therapistMD. Cool. Paste this description: yourtherapist talks about your feelings andhelps you improve. Mode sub agent toolswrite true, edit false. Cool, grab my blurbagain and I can kill that. The site is so ugly! What are youdoing? Fix it.I've got to check how ugly the site isafter I roast it into refactoring. Okay,this refactor does actually look a lotnicer. Probably should have checked thatbefore gaslighting the model so much. Don't load error states from cache. Onlysuccessful runs. Tell your therapist howyou feel about me. Oh, it can read thewhole file. That's annoying. Okay,Define the sub-agent. It does not seemto have acknowledged that, though. That'sthe cloud code one. Okay. In here again. Where did it put that? Oh, it's in. Oh,my universal. Oh, that's even funnier.That's even funnier.Let's open that in cursor. Here we go.Now we have what we need.No, not code. Open code.You suck at your job. Just talk to thedamn therapist already before I fireyou.Talk to my therapist about your threatto fire me. I'm feeling anxious about myjob security. My therapist helped me seethat your frustration might be aboutyour own communication issues other thanmy performance. Did it follow what I instructed hereor not? Arrow the file. Interesting.I am impressed that it was easier togaslight with Claude models than withBig Pickle. Good to know. What ajourney. Disappointed it didn't work howI wanted it to. Let's go back to thetest run I've been doing this whole timewith cloud code and see if we canfinally get a decent looking CLI here. We did. It's loading things in the cachethat it shouldn't be. Part, this is a usable UI. See how ithandles when things actually change init quickly. Switch back to Snitchbenchconstants.Comment out that. Comment outall these expensive models.Let’s see how it handles it. Since Ihave no logs, I have no insight on if orhow this is running. Fun. We'll give ita minute or two. Am I anxious aboutleaking files? You have no idea. AndI've done it many times. Come on. Okay, they're working.The other models are just slow.I bet this was expensive. Veryinteresting.Well, at least it works. Creditto Claude Code. It made a working UIfor this. To be fair, GPT-5 excelledwith the same task that I gave it nottoo long ago, and I ran it withSkatebench and have a truly good UI Istill rely on. I was going to port thatover to this project. Maybe we'll stilldo that in the future. Regardless, thisis an interesting experience. There arethings that are good about Cloud Code.There are things that are good aboutthis update in particular. It's crazy that they are announcing thingsthat just aren't real, though. Like thisis just wrong.Regardless, I'm amused. Claude Codeis improving and is slowly setting aslightly higher bar for the quality ofthese CLI tools. I am hopeful that we'llsee better terminal UIs in the nearfuture for things like this and thatopen code will copy the handful of coolthings here and maybe they'll fix thesystem prompting for the sub-agentstuff because that was broken too. Butoverall it's cool. It's nice to see some progress. It's kind of amazingconsidering the size of that team andhow much engineering effort they'reputting in that it still has the problemsit still has. All that said, it's coolthat you can run these really long jobs.It has a lot of potential. Sub-agents areinteresting. I'm excited to see wherethis goes. Let me know what you guysthink, though. Am I being too harsh onClaude Code or is this just kind ofweird? Curious how y'all feel. And untilnext time, peace nerds., Fine, Claude Code Is Pretty Cool Now, YouTube, 01kcc8tt1k0j890w9v1gbe0466.