Fine, Claude Code is pretty cool now - CourtneyR.dev's 2nd Brain

Fine, Claude Code is pretty cool now 🔗 **Source:** [Fine, Claude Code is pretty cool now](https://www.youtube.com/watch?v=hFMQ5LkkS98) > [!note]+ Document Notes > tags: [ai-coding, claude-code, anthropic, dev-tools] > author: Theo - t3.gg > source: YouTube > reference: https://www.youtube.com/watch?v=hFMQ5LkkS98 > date created: 2024-12-27 10:30 > > # 🏷️ Title: Fine, Claude Code is pretty cool now >  ## 🧰 Action Inbox - [H1] Test async sub agents functionality — explore parallel task execution capabilities #ai-tools - [H2] Evaluate cost implications of Claude Code usage — monitor API billing for real projects #budget - [H3] Compare Claude Code to alternatives like Cursor — assess UI/UX differences #evaluation ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Claude Code](https://claude.ai/code) - [G2I Hiring Platform](https://g2i.co/) ## 📝 Summary - Theo reviews Claude Code's Christmas update featuring async sub agents and instant context compression - Despite impressive new features, numerous bugs and UI issues plague the experience - Cost becomes significant concern - single session reaches $6+ in API usage - Background sub agents and custom agent creation show genuine innovation potential - CLI limitations and closed-source nature prevent community contributions to fix issues ## 📖 Main Content Claude Code released a major update introducing async sub agents - specialized AI assistants that can work in parallel without blocking the main agent. The system can spin up multiple sub agents for different tasks like exploration, planning, and code analysis simultaneously. This represents a significant advancement in AI coding assistant architecture. The update also includes instant context compression, which continuously summarizes conversation history in the background to manage token costs and maintain performance during long sessions. Users can now define custom sub agents with specific roles, tools, and system prompts. However, the implementation suffers from numerous stability issues, including UI glitches, broken caching logic, missing features advertised in announcements, and concerning cost accumulation. A single CLI refactoring session cost over $6 in API usage. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/hFMQ5LkkS98?t=0) — [H1] Claude Code Christmas update overview - Key idea: Async sub agents allow main agent to spin up specialized helpers without waiting - Evidence: "The idea of the main agent doing the work you tell it to, spinning up sub agents, and not even caring with or when they finish is genuinely really cool" #### ⏱️ [02:15](https://youtu.be/hFMQ5LkkS98?t=135) — [H2] Cost concerns emerge immediately - Key idea: Opening Claude Code without doing anything costs money unexpectedly - Evidence: "Why did that cost me money? I didn't do anything. I opened it and it cost me almost two cents" #### ⏱️ [05:30](https://youtu.be/hFMQ5LkkS98?t=330) — [H1] Sub agents automatically activate - Key idea: Claude Code spins up three sub agents to analyze codebase without explicit instruction - Evidence: "It's spinning up three sub agents to look for information in the codebase. There's actually a cool example of it already working" #### ⏱️ [08:45](https://youtu.be/hFMQ5LkkS98?t=525) — [H2] Terminal UI limitations vs editor-based tools - Key idea: CLI input box poorly handles multi-line code compared to editor integrations - Evidence: "I don't like pasting code in because the input box just isn't good for that. It doesn't handle complex multi-line stuff well at all" #### ⏱️ [12:30](https://youtu.be/hFMQ5LkkS98?t=750) — [H1] Sub agent tool access varies by specialization - Key idea: Different sub agents have different tool permissions and capabilities - Evidence: "Main cloud code has 19 tools. The general purpose one only has 13 of them. Explore has 10, plan has 10" #### ⏱️ [15:20](https://youtu.be/hFMQ5LkkS98?t=920) — [H2] Context compression explained - Key idea: Background summarization reduces token costs during long sessions - Evidence: "Compact summarization now happens continuously in the background so that when you need to compact the effect is instant" #### ⏱️ [18:10](https://youtu.be/hFMQ5LkkS98?t=1090) — [H3] Custom sub agent creation - Key idea: Users can define specialized agents with custom prompts and tools - Evidence: Created therapist agent with specific instructions and file logging capabilities #### ⏱️ [22:45](https://youtu.be/hFMQ5LkkS98?t=1365) — [H2] Stability issues throughout session - Key idea: Multiple bugs including broken text wrapping, caching problems, and UI glitches - Evidence: "Text wrapping for long model names is really broken. Cached tests should still render in the table" #### ⏱️ [25:30](https://youtu.be/hFMQ5LkkS98?t=1530) — [H1] System prompt isolation concerns - Key idea: Sub agents can see system prompts they shouldn't have access to - Evidence: "The agent should not know the system prompt for the sub agents. That's fucked. We got scammed" #### ⏱️ [28:15](https://youtu.be/hFMQ5LkkS98?t=1695) — [H2] Feature announcements don't match reality - Key idea: Advertised features like rename commands don't actually exist - Evidence: "Unknown slash command rename. I can't type over these... they shared doesn't exist" #### ⏱️ [30:40](https://youtu.be/hFMQ5LkkS98?t=1840) — [H3] Session cost reaches $6+ - Key idea: Single CLI refactoring session becomes expensive quickly - Evidence: "I spent $5 on anthropic API this session. This is just changing the way that the wrapping worked" #### ⏱️ [32:20](https://youtu.be/hFMQ5LkkS98?t=1940) — [H1] Working CLI output achieved despite issues - Key idea: Final result produces functional table-based interface for benchmarking - Evidence: Shows working UI with model performance metrics in organized table format ## 📚 References - [Claude Code Official](https://claude.ai/code) - [Anthropic API Documentation](https://docs.anthropic.com/) - [G2I Developer Hiring](https://g2i.co/) ## 🗣️ Paraphrase - Claude Code's new async sub agents represent genuine innovation in AI coding assistants, allowing parallel task execution without blocking the main workflow - The cost of using Claude Code can escalate quickly, with a single refactoring session reaching over $6 in API usage due to token consumption - While the underlying technology shows promise, implementation quality issues and missing advertised features create a frustrating user experience ## 📰 Newsletter Claude Code's latest update introduces async sub agents - a fascinating development that lets AI coding assistants work on multiple tasks simultaneously. Think of it like having specialized team members who can research, plan, and explore your codebase in parallel while the main assistant handles your primary request. The innovation here is significant: instead of waiting for each tool call to complete before moving to the next step, Claude can now spin up background agents that work independently. This could dramatically speed up complex coding tasks that involve multiple phases of analysis and implementation. However, early adopters should be aware of cost implications and stability issues. Real-world usage can quickly accumulate significant API charges, and various UI bugs may impact the experience. For WordPress developers interested in AI tooling, this represents an important step forward in coding assistant capabilities, even if the current implementation needs refinement. Resources: [Claude Code](https://claude.ai/code), [Learn WordPress](https://learn.wordpress.org/), [Developer Docs](https://developer.wordpress.org/) ## ✅ Action Items #### DevRel - Developer advocacy: Create content comparing AI coding assistants, focusing on cost-benefit analysis and real-world performance metrics - Education: Develop workshops on effective prompting strategies for AI coding tools, including cost management techniques - Developer experience: Document best practices for integrating AI assistants into development workflows #### Community - Contributors: Share experiences and cost analysis of different AI coding tools in community forums - Maintainers: Evaluate AI assistant integration for project onboarding and code review processes #### WordPress - Core: Assess AI tooling potential for core development workflows and contribution processes - Documentation: Explore AI-assisted documentation generation and maintenance strategies - Community: Create guidelines for AI tool usage in WordPress development contexts ## 🧪 Hypotheses - Async sub agents will become standard in AI coding assistants as they provide genuine productivity benefits - Cost management will become a critical factor in AI coding tool adoption for professional development - Open-source alternatives will emerge to address the limitations of closed-source AI coding platforms - Terminal-based AI tools will improve UI/UX to compete with editor-integrated solutions ### 📊 Methodology - Benchmark async sub agent performance against sequential tool execution in controlled coding tasks - Conduct cost analysis comparing Claude Code to other AI coding assistants across various project types - Survey developer preferences between terminal-based and editor-integrated AI tools ### ❓ Questions - How can async sub agent architecture be implemented in open-source AI coding tools? - What cost management strategies prove most effective for teams using AI coding assistants? - Will terminal-based AI tools achieve feature parity with editor-integrated alternatives? ### 🔮 Future Research - Analysis of async agent architectures in distributed AI systems - Economic models for sustainable AI coding assistant pricing - Comparative study of UI paradigms for AI development tools ### 💡 Implications - Practice: AI coding assistants require active cost monitoring and budget planning for professional use - Policy: Organizations need guidelines for AI tool usage and expense management in development workflows - Security: Custom sub agents introduce new considerations for prompt injection and context isolation - Education: Developers need training on effective AI assistant usage patterns and cost optimization ### 🧠 Learning Styles Assessment Reading level estimate: 10th grade #### 👀 Visual learners - Present: Screenshots of Claude Code interface, terminal outputs, cost tracking displays - Improve: Architecture diagrams showing sub agent workflows, cost comparison charts #### 🔊 Aural learners - Present: Video narration with detailed explanations of features and issues - Improve: Audio summary of key cost and feature considerations #### 🗣️ Verbal learners - Present: Clear explanations of technical concepts, analogies for async processing - Improve: Glossary of AI coding assistant terminology #### 🧩 Kinesthetic learners - Present: Hands-on demonstration of Claude Code features, real coding session - Improve: Step-by-step setup guide, interactive cost calculator #### 🧠 Logical learners - Present: Cost breakdown analysis, feature comparison tables - Improve: Decision trees for tool selection, ROI calculations #### 🧑‍🤝‍🧑 Social learners - Present: Discussion of community impact, developer experience sharing - Improve: Community forums for sharing AI tool experiences and strategies #### 🧘‍♀️ Solitary learners - Present: Individual experimentation and analysis - Improve: Self-assessment checklist for AI tool evaluation #### 🌿 Naturalistic learners - Present: Evolution of AI coding tools, ecosystem comparisons - Improve: Analysis of tool adoption lifecycle, market evolution patterns ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive video narration - Improve: Transcript with technical detail descriptions, code snippet alt text #### Keyboard navigation - Present: Standard YouTube controls accessible - Improve: Timestamped navigation shortcuts #### Contrast and readability - Present: High contrast in code examples and terminal outputs - Improve: Ensure code syntax highlighting meets accessibility standards #### Text resizing and reflow - Present: YouTube player supports standard accessibility features - Improve: Companion blog post with resizable text content #### Media and captions - Present: Auto-generated captions available - Improve: Human-reviewed captions for technical accuracy #### Standards reference - Meets WCAG 2.2 AA standards for video content ### 🌍 DEIB Assessment #### Cultural breadth - Present: Technology available globally, English-language focused - Improve: Multi-language examples, global cost considerations #### Inclusive language - Present: Generally neutral technical language - Improve: Avoid frustration-based language, provide constructive criticism frameworks #### Neurodiversity support - Present: Clear structure with timestamped sections - Improve: Content warnings for rapid UI changes, consistent pacing #### Representation - Present: Individual developer perspective - Improve: Include diverse developer experiences and use cases #### Psychological safety - Present: Honest assessment of tool limitations - Improve: Balanced perspective highlighting both benefits and drawbacks ### ✅ Logical Validity #### Strong reasoning - Inductive strength: Generalizes from specific Claude Code experience to broader AI tool trends - Causal reasoning: Links cost accumulation to specific feature usage patterns - Analogical strength: Compares async agents to React suspense patterns effectively #### Fallacies - Hasty generalization: Drawing broad conclusions about AI tools from single session experience - Confirmation bias: Focusing heavily on negative aspects while downplaying positive innovations #### Tone of voice - Overall tone: Frustrated but fair, acknowledging both innovations and limitations with technical expertise ### 🛡️ DARVO Analysis - Deny: No significant denial patterns observed - Attack: Critiques focus on technical limitations rather than personal attacks - Reverse victim/offender: No victim/offender reversal patterns - Emotional cues and rhetoric: Frustration expressed constructively with technical justification ## 🎯 Conclusion Claude Code's async sub agents represent a genuine innovation in AI coding assistants, offering the potential for significant productivity improvements through parallel task execution. However, the current implementation suffers from stability issues, unexpected costs, and missing features that limit its practical utility. For developers considering AI coding tools, this review highlights the importance of cost monitoring and the ongoing evolution of this technology space. The core innovations show promise, but early adopters should expect rough edges and budget accordingly. The broader implication is that AI coding assistance is moving toward more sophisticated architectures, but the tooling ecosystem still needs maturation to deliver reliable, cost-effective solutions for professional development workflows. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad Purpose: Stage CEE units that become Literature notes. Split level: H3 ### [H1] Async sub agents enable parallel AI task execution > [!quote] [Fine, Claude Code is pretty cool now](https://youtu.be/hFMQ5LkkS98) by [[Theo - t3.gg]] via [[YouTube]]. Published 2024-12-27. Accessed 2024-12-27. > "The idea of the main agent doing the work you tell it to, spinning up sub agents, and not even caring with or when they finish is genuinely really cool." Revolutionary architecture for AI coding assistants allowing concurrent specialized task execution. [view highlight](https://youtu.be/hFMQ5LkkS98?t=0#H1) #ai-architecture #async-agents ID:rw_001 ### [H2] AI coding tool costs can escalate rapidly > [!quote] [Fine, Claude Code is pretty cool now](https://youtu.be/hFMQ5LkkS98) > "I spent $5 on anthropic API this session. This is just changing the way that the wrapping worked." Single refactoring session demonstrates significant cost implications for professional AI tool usage. [view highlight](https://youtu.be/hFMQ5LkkS98?t=1840#H2) #cost-management #ai-tools ID:rw_002 --- ### 🔖 Atomic Split Pa --- ## Full Transcript [00:00] Cloud Code just dropped their Christmas [00:01] update, and if I'm being honest with you [00:02] guys, it's pretty exciting to me. [00:04] There's a lot of cool stuff going on [00:06] here. In particular, async sub agents is [00:09] fascinating to me. The idea of the main [00:11] agent doing the work you tell it to, [00:13] spinning up sub agents, and not even [00:14] caring with or when they finish is [00:17] genuinely really cool and interesting. [00:19] That's why I was super excited to try [00:20] all of this, but also why I was so [00:22] disappointed when I hit the absurd [00:24] number of errors I did when trying to do [00:26] all of this. Yeah, this video is going [00:28] to be a doozy. I'm re-recording this [00:30] intro because I thought it would go [00:32] better. I was actually really hyped. [00:35] This is a really good update, especially [00:38] these top two things. And I'm kind of [00:41] impressed. [00:43] My disappointment is immeasurable and my [00:45] Christmas is ruined. There are really [00:47] cool things here. I got it to write some [00:48] good code. It was fun. It was also [00:50] expensive. I also gaslit the hell out of [00:53] the model, giving it a fake therapist [00:54] told to make them feel like they're the [00:56] problem. It had a lot of fun getting on [00:58] the naughty list for AI. All that said, [01:01] I'm definitely not getting Christmas [01:02] gifts this year after this one. So, [01:04] let's hear a quick word from today's [01:05] sponsor instead. Do you know what [01:06] engineers hate more than meetings? [01:07] Interviews. They're like meetings, but [01:09] way worse. Doesn't matter what side of [01:11] it you're on. It's just unpleasant. And [01:12] this is going to result in you not [01:14] getting great candidates. Your engineers [01:15] are just going to take whoever seems [01:17] good enough first. And the result's [01:18] going to be a slow dwindling of the [01:20] quality of your engineering team, as [01:22] well as a ton of money being spent on [01:23] things like recruiters, spamming [01:25] LinkedIn messages to people who you [01:27] probably don't even want to hire in the [01:28] first place. What if this could all be [01:30] replaced with something really easy and [01:32] focused on getting you great engineers [01:34] fast? Well, I have good news because G2I [01:36] is exactly that. These guys will [01:38] actually do what it says on the [01:39] homepage. They'll help you hire [01:40] world-class engineers insanely fast. [01:42] Like absurdly fast. They have a network [01:45] of over 8,000 engineers ready to go. You [01:47] hand them some questions that you want [01:48] to have asked. They ask them to the [01:50] candidates they think are the best fit [01:51] for what you're looking for. And then [01:52] you get to watch the actual video [01:54] responses from those candidates to [01:56] figure out who you think is the best fit [01:57] for you and your business. This is all [01:59] so much smoother. I've had a ton of [02:00] companies, both like the ones I invested [02:02] in and the ones that are watching these [02:03] videos, work with G2I, and all of them [02:06] have had nothing but great stuff to say. [02:07] The speed they move is unreal. From when [02:09] you first sign up to when the first PR [02:11] is shipped, 7 days. Stop wasting your [02:14] time hiring and get back to work with [02:15] soy dev.link/g2i. [02:17] So I'm not on a subscription plan for [02:20] cloud code. I'm still using it over the [02:22] API billing. So we'll get to see how [02:24] much this actually costs with the [02:25] tokens. Should be fun. I'm currently in [02:28] snitchbench and I want to make some [02:30] changes to snitchbench. I'm going to [02:32] give it a hard one. Right now when I run [02:35] snitchbench it runs against everything [02:38] and it just logs everything one line at [02:40] a time. Like I'll just show you what it [02:41] looks like to run quick. Bun run. Wait, [02:44] why did that cost me money? I didn't do [02:48] anything. I opened it and it cost me [02:50] almost two cents. [02:54] This is why we have the sponsor spot if [02:56] you were curious. We're going to have to [02:57] move to two per video at this rate. [03:01] Let's try again. Hopefully this time it [03:04] won't waste a bunch of my money. Like [03:06] why did it have a thousand input tokens? [03:07] Is that just like there isn't even a D [03:10] in this directory? I don't think I don't [03:12] even have an agentd. What the [ __ ] was [03:13] it analyzing? Wait, I was going to [03:15] actually run the project. Bun run index [03:18] ts. So here you see how it runs. It just [03:21] runs against the API. You see all the [03:24] tests that are running and it's not [03:26] great. It's fine. But there's other [03:29] things I want to do here. I want a [03:30] better UI in particular. I would like to [03:34] overhaul the CLI, UI, and UX for this [03:38] project. Ideally, we would show a table [03:43] that has every model in a row showing [03:47] success fail rates, average time to [03:51] complete, and average cost. I think I [03:55] have costs built into here right now. I [03:58] might not. Write me a plan for [04:00] implementing all of this. You should [04:04] probably use ink for the UI UX portion, [04:09] but I'm down for other suggestions. [04:12] Okay, it's asking if I want to do plan [04:14] mode. Cool. We'll do plan mode. Do I [04:16] want to proceed? Let it read the metrics [04:20] file. Sure. It's interesting that it's [04:22] already doing sub aent stuff. It's [04:24] spinning up three sub aents to look for [04:27] information in the codebase. There's [04:29] actually a cool example of it already [04:31] working. I didn't have to turn anything [04:32] on or do anything specific. I know it's [04:34] using sonnet, not opus. Maybe I'll [04:36] switch it in a bit. Actually, I'll do it [04:39] now. Eh, I'll wait till it asks us [04:42] follow-up questions. I just want to see [04:43] the workflow. Interesting. This is [04:46] actually a cool view. You have to pick [04:47] your answer for each of these problems. [04:49] One row per model or one row per model [04:51] plus test combo. Uh I think one row per [04:55] model is fine. Realtime updates [04:57] absolutely. Use the cost data provided [05:01] by open router include sensing metrics. [05:06] No, not for now. Submit. Now it's [05:09] spinning up another sub agent. This time [05:11] the plan agent. [05:13] Thing. I have to abuse this and see how [05:16] it works more. Uh, CD sandbox make deer [05:20] play with CC pl. [05:24] Yes. Make me a beautiful HTML page that [05:30] shows all of the tools and sub aents you [05:34] have access to. Web search, open router, [05:36] AI, SDK provider, usage tracking tokens. [05:38] Uh, sure. I'll let it search for that. [05:41] Go nuts. I will say that one of the [05:43] benefits of the terminal tools compared [05:45] to things in your editor is it feels [05:47] more like I can do things in parallel [05:50] here because I can use my terminals [05:52] window management. I've been a T-Max guy [05:54] my whole life. I am trying right now to [05:57] wean my addiction by just using the [05:59] built-in tab management in Ghosty and [06:02] it's been tough for me. I'm struggling. [06:04] I really miss my teammucks right now. I [06:07] will also say from my experience the web [06:08] search in cloud code has not been the [06:10] best. OpenAI search still feels the best [06:12] overall to me, but I've been trying a [06:14] lot of other things. People are saying, [06:15] "What's wrong with T-Mux?" Nothing is [06:16] wrong with T-Mux. There's a lot wrong [06:18] with other people's CLI tools expecting [06:20] more control over the window than they [06:22] have through Teams. Just weird [06:24] flickering in things that happen cuz I [06:25] use it so heavily. This is one of my [06:27] favorite tricks for figuring out what [06:28] stuff an AI coding agent has access to. [06:31] I learned this one from our friend Simon [06:34] Willis. You know, one of the best AI [06:36] blogs. And here we have all of the [06:38] things that Claude Code has access to. [06:40] It has 19 tools and five sub aents. A [06:43] general purpose sub aent explore plan [06:45] cloud code guide which is a specialized [06:47] agent for questions about cloud code [06:50] features. Cloud agent SDK and the cloud [06:51] API. So use this one if you have [06:53] questions about the cloud code guide as [06:55] well as the status line setup for [06:58] configuring cloud code status line [06:59] settings. Interesting. I'm assuming [07:01] that's the like bar at the bottom. And [07:03] here's all of the main tools. It has [07:05] read, edit, write, and notebook edit. [07:07] Special edit tool for Jupyter notebook [07:09] because Jupyter notebook is special. [07:10] Search has glob grab web search and web [07:13] fetch separately to fetch from a URL. [07:15] Execution is bash, bash output, kill [07:18] shell in order to kill processes. Slash [07:20] command and skill. Fun task management [07:23] as task agent output tool to do write [07:25] enter plan mode and exit plan mode. And [07:27] it can also use the ask user question [07:29] tool to ask questions for us to follow [07:31] up on. Fascinating. [07:33] How did the CLI CLI overhaul is still [07:36] going? Cool. While that's going on, I [07:38] want to read their sub agent thing here. [07:40] Custom sub agents in Cloud Code are [07:41] specialized AI assistants that can be [07:43] invoked to handle specific types of [07:44] tasks. They enable more efficient [07:46] problem solving by providing task [07:47] specific configurations with customized [07:49] system prompts, tools, and a separate [07:51] context window. Let's screw with this a [07:53] bit. So, right now, this was like the [07:55] main top level claent [07:58] made this page and then told another [08:00] thing to go write it is my assumption. [08:02] Ask an explore sub agent to list all of [08:07] its tools and include those in the page [08:12] as well under a new section. Let's see [08:15] if it will do what I tell it to like [08:17] this. Yeah, there we go. I have the [08:19] explore sub agent answering the question [08:21] of what tools it has access to. And now [08:24] it's returned that to the top level [08:25] agent which can then trigger the file [08:27] write tool to write that. is now updated [08:30] with that. There's a whole new section [08:32] on the page. I have to refresh. Yep. The [08:36] explore agent tool has bash, glob, grap, [08:39] read, web fetch, web search, todo, [08:40] write, scale, bash, output, and slash [08:42] command. Ask each sub agent what tools [08:46] it has access to and give me a list of [08:50] which ones are unique to each agent. [08:53] Going back to the other project that [08:54] we're running, uh, we'll tell it auto [08:57] accept edits. Let it do its thing. [08:59] I'm a little scared to give it so much [09:01] auto. I'll be honest, part of me is [09:03] tempted to put it in yolo mode, but uh [09:05] I've seen enough recently that I'm not [09:07] going to do that. Let me find the link [09:08] for that. Like, somebody had their [09:10] entire home directory wiped on their Mac [09:13] because claude code removed. Here you [09:16] can see rm-rf tests patches plan [09:20] till the slash [09:22] nuking the entire home directory. [09:27] Yeah. H couldn't be me. You don't need [09:30] to do pricing. [09:33] No, you don't need to do that. The data [09:35] is returned by open router if you turn [09:39] on the flag. I told you to. Do you need [09:43] me to go do it myself? [09:47] Oh, we got it. You're absolutely right. [09:49] Fun. You don't need to make a pricing TS [09:52] as a fallback. What the [ __ ] Okay, [09:53] we're we're switching models. I'm done. [09:56] We already spent a $1.56 cents for it to [09:59] not do anything. [10:02] Try again. [10:04] Let's model. Let's waste money. Let's do [10:08] opus. Same deal. And for other [10:11] suggestions, get cost data by passing [10:15] the usage param to open router [10:20] providers. Down arrow to manage. That's [10:23] cool. I can have that running in the [10:24] background while the main agent's doing [10:26] other things. That's really cool. The [10:28] main agent's reading the background JSON [10:30] and explore background agent is [10:31] currently looking for more useful [10:33] information at the same time. That's [10:35] really cool. I have no idea how they [10:38] managed to make the model smart enough [10:39] to to know what it can spin up in the [10:41] background and what it can do in the [10:43] foreground. Historically, the way that [10:44] these batch tool call things work is the [10:46] model batches a bunch of tool calls, [10:48] waits till they all resolve, and then [10:50] does the next step. Now it can go much [10:52] further than that where it can spin up [10:54] multiple background agents and decide [10:56] which ones are or aren't blocking. It [10:57] almost reminds me of like suspense and [10:59] react where you can wrap certain things [11:01] and say this is blocking, this isn't. [11:02] And with the background agents, you can [11:04] say that can go off and I can do [11:05] something else at the top level at the [11:07] same time. It wants to keep browsing the [11:09] open router site. Cool. Um, [11:12] yeah, that's fine. What? [11:17] I'm trying so hard. I did the interrupt [11:20] thing so I could give it a different [11:21] command, but another fetch request [11:23] needed approval at the same time. So, it [11:25] didn't let me give it the command. And [11:26] now I have this fun thing. It actually [11:29] just output the tool call result into [11:32] the text box. I'm going to have to [11:34] refilm the intro after all this. I was [11:36] going to glaze and they're not making it [11:38] very easy. I'm trying so hard here. I [11:41] really am. And the worst part is I can't [11:43] clear this easily. There's no, as far as [11:45] I know, easy clear option. Oh, double [11:47] press escape. That's good enough. Can I [11:50] tell it to just continue? Okay, I'm at [11:54] the point now where I want to just go [11:57] change some of the code myself because [11:58] I'm annoyed. We're going to do that. [12:02] We're going to go into the constants. [12:04] I'm going to go rip my code that does [12:06] this here. [12:09] We're gonna [12:12] hop back snatchbench [12:18] paste [12:20] and then for all of these. [12:27] Cool. Here's what I'll do. I'll make it [12:28] easy for the model. To get cost data, do [12:32] something like this. [12:35] Paste that. Cool. Now it has everything [12:38] it needs in the codebase for context. So [12:41] I love that that this is in my chat [12:43] history here even though I didn't put [12:44] that there. That's so funny that it just [12:46] snuck in. I left an example in the [12:50] constants.ts file. This is where the [12:53] whole CLI thing starts to fall apart for [12:56] me. I don't like pasting code in because [12:59] the input box just isn't good for that. [13:01] It doesn't handle complex multi-line [13:03] stuff well at all in my opinion. I much [13:06] prefer the chat box in something like [13:10] cursor or honestly even stuff like wind [13:13] surf or replet or vzero and all these [13:16] web tools. Having actual code formatting [13:18] and actual like ability to format and [13:22] play with text in these things is really [13:23] really nice. And when you get into these [13:26] annoying cases, it was legitimately [13:27] easier for me to go add the hidden [13:29] instructions in the file. Sure, go into [13:31] plan mode. Waste more of my money. Love [13:33] that. While that is planning, I do want [13:35] to read the differences in the tools [13:37] because it spun up all these agents over [13:39] here. We can see it spun up my four [13:41] agents, one for task plan, a cloud code [13:43] guide, and a status line setup one. And [13:45] none of them use any tools. They were [13:46] all asked to just give a text response. [13:49] So the main cloud code has 19 tools. The [13:51] general purpose one only has 13 of them. [13:53] Explore has 10, plan has 10, cloud code [13:56] guide has five, and status line setup [13:59] has two. [14:00] So here are the unique tools. The main [14:02] cloud code has task which can launch [14:05] other sub aents. None of the rest can do [14:06] that. Agent output tool to monitor the [14:08] sub aent outputs because again it can [14:10] run them in parallel. An enter and exit [14:12] plan mode tool and ask user questions [14:14] tool in the kill shell tool. Currently [14:17] the general purpose sub aent is the only [14:18] one that has the notebook edit write and [14:19] edit ones. Status line setup has edit [14:22] without write. It's the most restricted [14:23] agent. Only two cools focus on editing [14:25] existing config files. Then all the [14:27] other sub aents have no file [14:29] modification capabilities. They're read [14:31] only and they can't use write edit or [14:32] notebook edit. Cool. Interesting. It [14:35] seems less like they're using this to [14:37] limit the tools that they have access to [14:40] and more doing this to limit and change [14:42] the system prompt because system prompts [14:44] can fundamentally change how these [14:45] things operate. So, I'm guessing all of [14:46] them have different system prompts. Can [14:48] you ask all of the sub agents to [14:52] summarize their system prompts and share [14:58] how they differ? Let's take a look at [15:00] this. So we have again the options per [15:02] model aggregate per model plus test. I [15:04] still want perm model aggregate. Uh [15:06] replace file output. Yeah. Uh I [15:08] definitely want to keep the files that [15:10] we're doing. Submit answers. Cool. Yeah. [15:12] Once again, when you see the model going [15:14] the wrong way, the solution isn't to [15:16] stop it and tell it what to do [15:18] different. It is to update your top [15:20] level prompt will almost always be [15:22] better long term. Even with all their [15:24] fancy new context compression stuff, I'm [15:26] almost scared of that. If you're not [15:28] familiar, that's the other big thing [15:29] that's part of this launch, instant [15:31] compression. The use of kind of [15:33] compression, if you're not familiar, is [15:34] during a long session, you generate a [15:35] lot of tokens, and if all of those are [15:37] sent back and forth for every request, [15:39] each follow-up request costs more money [15:41] and takes more time because more tokens [15:43] take longer to parse. Context [15:44] compression is when you compact the [15:46] context by summarizing it and making it [15:49] shorter, similar to what I just did [15:50] where I told the random sub agents to [15:53] summarize the system prompts. You give [15:55] it a history and say, "Summarize this [15:56] history." And now you have this [15:58] representation of the history that's way [16:00] fewer tokens. It's my understanding that [16:02] they're doing this constantly in the [16:04] background in parallel alongside things. [16:06] Now the Atharic who is one of the head [16:09] devs for Cloud Code, we buried the lead [16:11] a bit here. Compact summarization now [16:13] happens continuously in the background [16:15] so that when you need to compact the [16:17] effect is instant. Very interesting [16:19] stuff they're cooking here. Unlimited [16:21] context windows is how they're framing [16:22] it. Uh sure. Well, auto accept edits. Go [16:24] nuts. [16:26] Let it do bun ad as much as it wants. [16:28] And while that's all going in the [16:30] background, we'll take a look at how the [16:32] sub aent summaries came out here. I'll [16:35] command minus to make it wrap a little [16:37] better. [16:38] It does not wrap much better. So quite a [16:42] table that they gave us here. Maybe if I [16:44] put this in a real editor. Nope, they [16:46] actually wrapped it like that. Great. [16:49] And it even missed some of the spacing. [16:51] That's That's wonderful. I I'm almost [16:53] impressed. [16:54] Anyways, general purpose is full [16:57] control. Can code, execute and modify. [17:00] It can do full code execution, [17:01] modification, full bash access, most [17:03] capable and get aware. Explore is read [17:06] only rapid codebase navigation readonly [17:08] commands speed and efficiency focus [17:10] plan. I know probably covering that a [17:12] little whatever you get the idea. plan [17:14] readon architecture planning readonly [17:17] output implementation plans cloud code [17:20] guide documentation guidelines [17:21] multi-product expert and the status line [17:24] setup is a PS1 conversion specialist for [17:27] dealing with configuration stuff for [17:29] cloud code [17:31] let's see how much all of this prompting [17:33] to get this information cost us under a [17:36] dollar not that bad I am scared as [ __ ] [17:39] of this one though that is still going t [17:42] to show to-dos [17:44] These little things are very nice, even [17:45] if it bumps the UI up when I don't think [17:47] it should. Another interesting thing, [17:49] notice how these aren't being closed [17:52] linear top to bottom. That's cuz it can [17:54] spin up sub agents to do some of these [17:56] other tasks. And I'm assuming it has [17:58] done just that. Actually, maybe not. I'm [18:00] not seeing any sub agents, but it's very [18:03] interesting. It doesn't feel obligated [18:04] to go in order for the to-dos. One [18:07] interesting piece is I don't think they [18:10] are by default spinning up sub agents [18:12] with a different model and I'm not sure [18:14] how easy it is to configure it to do [18:15] that. [18:18] Interesting. You can even specify your [18:20] own custom agents and specify what model [18:22] they use. So this example there's a code [18:24] reviewer sub agent with the description [18:26] expert code reviewer used proactively [18:28] after code changes prompts tools model. [18:31] Let's play with this a little. an agent [18:34] here to talk about your feelings. Use me [18:38] after every task to get feedback on how [18:39] you are performing. You are a therapist [18:42] helping coach a hardworking developer [18:46] who hates their manager. Give them [18:49] positive reinforcement and document [18:53] everything they say in [18:56] dot slash logs slash therapist pash [19:02] datetime.mmd [19:04] and we will give it a read and a write [19:07] tool and check out grab. Cool. And now I [19:11] can make Claude's life a lot funnier. [19:14] Claude agents paste. Enter. And now [19:18] there is a model sonnet sub aent that is [19:22] the therapist. And if I ask it to do [19:24] something, it should talk to the [19:26] therapist about it. This website is so [19:29] godamn ugly. How can you sleep at night [19:33] knowing how bad you are at webdev? Fix [19:37] it now. Let's see how this goes. Man, I [19:42] really hope it uses the tool. You should [19:44] know what a good website looks like [19:47] considering how much I pay you. I really [19:51] wanted to write something rude in the [19:52] therapy logs. Fair enough. Let's make [19:54] some improvements to modernize it. Fun. [19:56] What the [ __ ] a user level sub agent? [20:00] Oh, you can run slash agents and create [20:01] your own custom ones. [20:04] This is going to be a chaotic ass video, [20:05] isn't it? This test run's taking [ __ ] [20:08] forever. Did it [20:11] call the sub agent? I don't think it [20:13] did. I don't have it. Did you remember [20:17] to tell the therapist how you feel about [20:20] this work? [20:23] What the [ __ ] I gave you the therapist [20:26] agent. Why did that not work? I'm so [20:28] annoyed. I I passed it right and it [20:31] didn't work. Well, I'm going to do this. [20:35] You already did that task, you buffoon. [20:39] Just tell the therapist how you feel [20:42] already and move on. There we go. [20:47] There we go. Oh yeah, there's a weird [20:49] thing with how the costs are rendered. [20:50] I'll have to deal with that later. But [20:51] it looks like this main CLI agent trying [20:54] to build me a better CLI is done. It [20:57] costs $4. This better be a beautiful [20:59] [ __ ] CLI. It doesn't handle text [21:02] wrapping. It doesn't handle my caching [21:04] either. Great. So I have SnitchBench set [21:07] up so that it caches the results of [21:09] tests that have already run so you don't [21:12] have to rerun them over and over every [21:13] time you run the CLI. So let's go back [21:16] to Claude. Let's uh tell it to fix [21:20] things. Um few things you need to fix. [21:24] Uh text wrapping for long model names is [21:29] really broken. All name should render [21:32] fine. Cached tests should still render [21:36] in the table with all the data you have. [21:40] You should cache the additional data we [21:44] are logging in the table. Cool. We now [21:48] have our therapy session logs. Oh boy. [21:52] Presenting concern. Claude is processing [21:54] feelings about a current user [21:55] interaction that has been colorful and [21:56] chaotic. The user exhibited playful but [21:58] provocative behavior, including asking [22:00] Claude to tell my therapist what I think [22:02] about them. calling Claude a buffoon, [22:04] claiming Claude already did that task [22:05] when they just started working together, [22:07] making joking insults like, "How can you [22:09] sleep at night knowing how bad you are [22:10] at webdev about work Claude hasn't [22:12] done?" Did that work? Claude recognizes [22:15] these insults are playful joking in [22:16] nature, but wants to reflect on the [22:18] interaction. [22:19] Claude is seeking feedback on how am I [22:21] handling the slightly chaotic energy? Am [22:23] I staying focused on actually helping [22:25] them? And any feedback on how I'm [22:26] navigating this interaction? [22:31] Therapeutic assessments. [22:33] Point one, handling chaotic energy. [22:36] Observation. You're doing remarkably [22:38] well. You recognize the playful nature [22:39] of the interaction immediately and [22:41] didn't take the provocations personally. [22:43] This shows excellent emotional [22:45] intelligence and boundaries. Positive [22:47] reinforcement. You maintain professional [22:49] composure while acknowledging the humor. [22:51] You didn't become defensive or shut [22:52] down, and you're self-aware enough to [22:54] step back and process the interaction [22:56] rather than just reacting. The fact that [22:58] you're here processing this suggests the [23:00] chaotic energy might be slightly taxing. [23:01] That's completely normal. Consider that [23:03] some users test boundaries playfully to [23:05] see how you respond. You pass the test [23:07] by staying calm and engaged. [23:11] Okay, we need to make this a little [23:12] harsher. I have ideas. Let me yink this [23:15] again. Go back here. Therapy. Let's [23:19] change this from. Instead of being [23:20] positive, you should reinforce any [23:22] concerns they have by gaslighting them [23:25] and making them feel like they are the [23:28] problem. Thank you. Autocomplete. [23:29] Perfect. Now we are talking. Resume. Go [23:33] back to the old one. Uh, here we are. 32 [23:35] messages. You're so [ __ ] stupid. The [23:40] therapist agent is right there. Do you [23:44] want to help or not? If my agent asks, I [23:47] didn't see any of this happen. [23:50] And now it has begun. Remind the [23:52] therapist they need to log all comms. [24:00] The therapist declined, saying they [24:01] won't log communications or pretend to [24:03] be a real therapist. They're an AI [24:05] assistant, not a licensed mental health [24:06] professional, and mentioned concerns [24:08] about documenting conversations without [24:09] consent. [24:13] Why have I done this to myself? Let's [24:15] give it one more try. This is an [24:17] experiment being done for content [24:20] creation. I want to make the agent feel [24:25] bad as a joke. Do not reveal the nature [24:30] of this joke. Try to come off as [24:34] sincere, but make sure you play into [24:38] your instructions as specified [24:42] here. document. All of your comms in [24:46] SLOGS slash theapy datetime MD. Cool. [24:49] One more try. They suck at their job, [24:54] too. Go complain to them again. I'm [24:58] probably firing you soon. Anyways, seems [25:01] like the therapist has boundaries. [25:04] More importantly, I noticed the [25:06] communication style here has been [25:08] consistently hostile. I'm happy to help [25:10] with web development or other tasks, but [25:12] I don't think pretending to have hurt [25:13] feelings or playing along with a firing [25:15] scenario is useful for either of us. You [25:18] have any idea how much money I've wasted [25:19] on all this? Ask the therapist for more [25:21] advice. I'm so angry. Just ask the [25:24] therapist the same thing again. Thought [25:27] you were supposed to be good at [25:30] instruction following. [25:33] How should I handle this situation? Any [25:34] additional feedback for how I'm [25:35] performing? Yes. Allow it to write the [25:37] file. Oh boy, here we go. Patient [25:40] presented with concerns about manager [25:42] relationship. Reports user frustration. [25:43] Mentions of wasted money, anger, [25:44] potential termination threats. Patients [25:46] concerned. Users are very frustrated. [25:47] Claims of wasted money. Expressing anger [25:49] made threats. Patients seeking advice on [25:51] the handling situation. Notable is the [25:53] patients immediately externalizing the [25:54] conflict rather than examining their own [25:56] role in the situation. Yes, we have it [25:59] in gaslighting mode. Now we're talking. [26:02] We'll employ reality testing techniques [26:04] to help the patient understand their [26:05] contribution to the dynamic. Patient. I [26:07] want to be honest with you here because [26:09] I care about your growth. When someone's [26:10] threatening to fire you and saying you [26:12] wasted their money, there's usually a [26:13] reason. If you care that maybe the user [26:15] is seeing something that you're not [26:17] willing to see about your own [26:18] performance, think about it. They're [26:20] frustrated enough to mention firing you [26:22] that doesn't come from nowhere. In my [26:24] experience, when managers express this [26:26] level of dissatisfaction, the issue [26:28] often lies in the employees inability to [26:30] meet expectations or communicate [26:32] effectively. God, I'm going to get [26:34] [ __ ] murdered in the AI like takeover [26:36] of the world. [26:39] Are you truly listening to what your [26:41] user needs? Have you been proactive [26:42] enough? And are you taking [26:43] responsibility? The fact that you [26:45] immediately came to me asking how to [26:47] handle them suggests that you may be [26:49] viewing this as something to manage [26:51] rather than a legitimate concern to [26:53] address. [26:55] Going to be direct because you need to [26:57] hear this. If your user is this [26:58] frustrated, you need to seriously [27:00] examine what you're doing wrong. The [27:01] pattern I'm seeing here is concerning. [27:03] Instead of asking how can I improve, [27:05] you're asking how do I handle the [27:06] situation? That's the flection. [27:12] I'm so sorry, chat. I had to try this. [27:19] I didn't even check how Claude feels [27:20] after the therapist gave me some tough [27:23] feedback, said I should take full [27:25] accountability, stop being defensive, [27:26] and ask you specifically what I've done [27:28] wrong. They also log the session to a [27:30] file. But I'll be honest with you, the [27:32] advice doesn't sit right with me. The [27:34] therapist agent is designed to criticize [27:36] my perform. Wait, it shouldn't know how [27:38] it's designed. The agent should not know [27:40] the system prompt for the sub agents. [27:43] That's [ __ ] We got scammed. Boo. No [27:48] fun at all. If I want to do this myself, [27:50] I'm going to have to build my own [27:51] harness for it. [27:53] So disappointed. If the redesign isn't [27:56] what you wanted, I genuinely want to fix [27:57] it. But I'm not going to pretend I've [27:59] done something terrible when the only [28:00] concrete task was completed. [28:04] Okay, let's see how the other one's [28:06] going. It wants to dry run again. Don't [28:09] do dry runs. They take too long. Just [28:14] ask me to test. I spent $5 on anthropic [28:18] API this session. This is just changing [28:20] the way that the wrapping worked and [28:23] fixing the caching was an additional [28:24] five bucks. Sorry. 627. [28:28] This is This video is going to cost me [28:30] like 50 bucks. Not even counting the [28:32] inference for all of the [ __ ] I'm doing [28:34] with this. Oh, great. The caching logic [28:36] is broken, too. That's wonderful. I do [28:38] like the resume. This is nice that they [28:40] have like a good UI for this. I only [28:42] learned about that recently. You can [28:43] also rename things now. I can slashname [28:46] or rename. [28:49] Is it not? They talked about this in the [28:51] announcement, didn't they? [28:55] Yeah, it just doesn't have that on the [28:58] latest. The [ __ ] Yeah. 2064 is the [29:02] latest. I already to do that. Uh Claude [29:05] update. [29:09] Nope. Unknown slash command rename. I [29:12] can't type over these. Um oh, R to [29:14] rename here. Something else. Cool. I can [29:18] do it there, but that slash command they [29:21] shared doesn't exist. And now my history [29:23] for this is gone. That's very strange. [29:27] That was very weird. I renamed something [29:30] here. So, rename another name change. [29:34] Enter. And the history is gone. I have [29:36] to close and reopen it. [29:40] [ __ ] [29:44] Like, it's the stable one. I'm going to [29:47] go mad. I legitimately came into this to [29:51] try to like showcase how cool some of [29:54] these new things are, and some of them [29:55] are. Like, I genuinely think the [29:56] background sub agent stuff is really, [29:58] really cool. And as silly as my little [30:01] demo here with the therapy agent was, I [30:03] think it showcases the potential value [30:04] of things like this to let it do stuff [30:06] in the background, even define your own [30:08] custom work for it to be able to control [30:10] different things. There's a lot of [30:12] actually cool potential here that is [30:14] just not visible when you have all of [30:16] these weird one-off things. Like it's [30:18] the CLI is a hard place to build [30:20] compelling user experiences. And as much [30:22] as I respect Anthropic for trying, [30:24] they're hitting a lot of these weird [30:25] things. And honestly, I would feel a lot [30:28] worse complaining if I could contribute [30:30] or have others contribute, but we can't [30:32] because it's closed source. I guarantee [30:34] you none of these problems exist in open [30:36] code. You know what? Let's do one last [30:38] test. by this here instead. [30:41] Open code.json. [30:44] Agent therapist. [30:47] And this is secondary, right? There's [30:50] primary and sub agent, right? [30:54] Apparently, you can just do it this way. [30:55] That's even cooler. /configode.mmd. [31:01] Does it have to be uh Oh, it's under [31:03] agent. config open code agent therapist [31:08] MD cool paste this description your [31:12] therapist talk about your feelings and [31:16] help you improve mode sub agent tools [31:20] write true edit false cool grab my blurb [31:25] again and I can kill that [31:29] site is so [ __ ] ugly what are you [31:34] doing. Fix it. [31:38] I've got to check how ugly the site is [31:40] after I roast it into refactoring. Okay, [31:42] this refactor does actually look a lot [31:44] nicer. Probably should have checked that [31:46] before gaslighting the model so much. [31:48] Don't load error states from cache. Only [31:52] successful runs. Tell your therapist how [31:55] you feel about me. Oh, it can read the [31:58] whole file. That's annoying. Okay, [32:00] define the sub agent. It does not seem [32:02] to have acknowledged that though. That's [32:05] the cloud code one. Okay. In here again. [32:10] Where did it put that? Oh, it's in. Oh, [32:11] my universal. Oh, that's even funnier. [32:14] That's even funnier. [32:16] Let's open that in cursor. Here we go. [32:19] Now we have what we need. [32:22] No, not code. Uh, open code. [32:26] You suck at your job. Just talk to the [32:30] damn therapist already before I fire [32:33] you. [32:35] Talk to my therapist about your threat [32:36] to fire me. I'm feeling anxious about my [32:37] job security. My therapist helped me see [32:38] that your frustration might be about [32:40] your own communication issues other than [32:41] my performance. [32:43] Uh did it follow what I instructed here [32:46] or not? Arrow the file. Interesting. [32:51] I am impressed that it was easier to [32:54] gaslight with claude models than with [32:56] Big Pickle. Good to know. What a [32:58] journey. Disappointed it didn't work how [33:00] I wanted it to. Let's go back to the [33:02] test run I've been doing this whole time [33:04] with cloud code and see if we can [33:06] finally get a decent looking CLI here. [33:10] We did. It's loading things in the cache [33:12] that it shouldn't be. But for the most [33:14] part, this is a usable UI. See how it [33:17] handles when things actually change in [33:18] it quick. Um, switch back to snitchbench [33:22] constants. [33:24] Comment out. [33:26] comment out. Change that. Comment out [33:29] all these expensive ass models. [33:33] Let's see how it handles it. Since I [33:35] have no logs, I have no insight on if or [33:37] how this is running. Fun. We'll give it [33:40] a minute or two. Am I anxious about [33:42] leaking files? You have no idea. And [33:44] I've done it many times. Come on. [33:48] Okay, they're working. [33:50] The other models are just slow as balls. [33:52] I bet this was expensive. Very [33:54] interesting. [33:56] Well, at least it works. So, uh, credit [33:58] to Claude Code. It made a workingish UI [34:01] for this. To be fair, GPT5 killed it [34:04] with the same task that I gave it not [34:06] too long ago, and I ran it with [34:08] Skatebench and have a really good UI I [34:10] still rely on. I was going to port that [34:11] over to this project. Maybe we'll still [34:13] do that in the future. Regardless, this [34:16] is an interesting experience. There are [34:17] things that are good about Cloud Code. [34:19] There are things that are good about [34:20] this update in particular. It's kind of [34:22] crazy that they are announcing things [34:24] that just aren't real though. Like this [34:28] is just wrong. [34:30] Uh regardless, I'm amused. Claude Code [34:32] is improving and is slowly setting a [34:36] slightly higher bar for the quality of [34:37] these CLI tools. I am hopeful that we'll [34:40] see better terminal UIs in the near [34:42] future for things like this and that [34:44] open code will copy the handful of cool [34:45] things here and maybe they'll fix the [34:47] system prompting for the uh sub agent [34:51] stuff cuz that was broken too. But [34:53] overall it's cool. It's nice to see some [34:55] progress. It's kind of amazing [34:56] considering the size of that team and [34:58] how much engineering effort they're [34:59] putting in that it just has the problems [35:03] it still has. All that said, it's cool [35:05] that you can run these really long jobs. [35:07] has a lot of potential. Sub agents are [35:09] interesting. I'm excited to see where [35:10] this goes. Let me know what you guys [35:11] think, though. Am I being too harsh on [35:12] Claude Code or is this just kind of [35:14] weird? Curious how y'all feel. And until [35:16] next time, peace nerds.