The argument in one line.
Getting an agent to work for long stretches unsupervised comes down to three tools — /goal for exit conditions, workflows for splitting worker and verifier sub-agents, and a planning habit centered on surfacing unknowns before writing any spec.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code and want to move past one-shot prompting into longer, more autonomous runs.
- You're a non-technical creator (video, content, marketing) who wants a repeatable Claude-driven production pipeline.
- You're coordinating several parallel Claude sessions at once and hitting context-window fatigue or thread overload.
- You write specs or PRDs and are wondering how that format needs to change now that an agent reads it too.
SKIP IF…
- You want an exact command-syntax tutorial — this is a conceptual conversation between two Claude Code power users, not a step-by-step walkthrough.
- You want the Remotion/Whisper code itself explained line by line — the demo is discussed and shown, not built on screen from scratch.
The full version, fast.
Anthropic's Thariq Shihipar breaks down three tools for getting Claude to run longer unsupervised: /goal locks in an exit condition so the agent won't stop early on a scoped task, workflows spin up separate worker and verifier sub-agents (because a model grading its own output is lenient — "self-preferential bias"), and planning itself is reframed as an iterative process of surfacing unknowns, not a one-time handoff. He live-demos a one-prompt video pipeline (Whisper transcription plus Remotion overlays) and an HTML-artifact design-exploration tool, describes running parallel Claude sessions through Slack ("Claude Tag"), and explains why Anthropic cut Claude Code's system prompt by 80% as the models got smarter and needed fewer examples and hard constraints.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 04:48
01 · Goals, loops, and workflows to make Claude work longer
Thariq frames /goal, /loop, and workflows as three related ways to get an agent to work for long stretches in an orchestrated way — /goal locks in the exit condition, workflows add parallel sub-agents that both do and verify the work.
04:48 – 08:17
02 · Live demo: Creating a polished video with one prompt
Thariq shows a video he recorded minutes earlier, then a single Claude Code prompt that transcribes it with Whisper, builds word-highlighted captions and overlays with Remotion, and renders straight through to a finished fade-to-black — one shot, no back-and-forth.
08:17 – 14:32
03 · Plan by removing unknowns before you build
Thariq reframes planning as an iterative process of discovering unknowns rather than a one-time handoff, using a Claude-generated explainer on how Whisper actually works (and its specific failure modes) as an example of research that builds confidence before committing to a workflow.
14:32 – 18:35
04 · How Anthropic uses HTML artifacts for planning
Thariq generates an HTML artifact exploring different overlay/caption design directions styled after Peter's blog, and explains that HTML artifacts have become Anthropic's default shareable format for plans, PRs, and status reports internally.
18:35 – 27:34
05 · Running parallel Claude tasks through Slack
Thariq describes 'multiclouding' — running most background and parallel work (PR babysitting, initial exploration, specs) through Claude Tag in Slack, where each channel keeps its own memory, while keeping one focused active session in Claude Code.
27:34 – 31:08
06 · Using separate creator and verifier sub-agents
Discussing Peter's short-form clip pipeline, Thariq explains why workflows split a worker sub-agent from a separate verifier sub-agent grading against a rubric — because a model judging its own work exhibits self-preferential bias and is too lenient.
31:08 – 34:07
07 · How Thariq stays focused while Claude works in parallel
Asked how to avoid burnout running five threads at once, Thariq says his 2026 goal is to be more productive but work less, and the habit that actually saves time is picking one project to focus on rather than half-supervising several at once.
34:07 – 37:01
08 · Why Anthropic cut Claude Code's system prompt by 80%
Thariq explains that as models got smarter they needed fewer directions, constraints, and examples — heavy example lists and blanket 'never do this' rules end up over-constraining behavior, so the team trimmed the prompt and leans on explaining the reason behind a constraint instead.
37:01 – 41:17
09 · How to become more technical when AI writes the code
Closing out, Thariq argues becoming more technical isn't about learning syntax but about learning a system's tradeoffs and constraints well enough to spot your own unknown unknowns — and that takes deliberately pushing Claude to explain itself, not just accepting its output.
Lines worth screenshotting.
- /goal locks in an agent's exit condition so it won't stop early on a task you've already fully scoped — it's built for handing off execution, not for exploration.
- A model grading its own output is measurably more lenient than a separate agent grading it — Anthropic calls this self-preferential bias, and it's the core argument for splitting worker and verifier sub-agents.
- Workflows are worth the overhead specifically when 'good' isn't a deterministic check — a rendered UI matching a Figma file can be verified directly; a short video's quality needs a rubric and a separate judge.
- Giving an agent a Figma file to verify against is easier than a screenshot, because the Figma MCP gives it a checkable spec instead of a fuzzy visual match.
- Planning with an agent is not a one-time handoff — it's an iterative loop of exploring, building small, and re-specifying once implementation surfaces things nobody anticipated.
- The real goal of a planning pass is finding your unknown unknowns, not writing everything down once — technical research, mockups, and explainers all count as planning.
- Anthropic cut the Claude Code system prompt by 80% because smarter models need fewer examples and fewer constraints — heavy examples end up over-constraining behavior instead of just illustrating it.
- A blanket 'never do X' instruction is usually a lie — you rarely mean never, and explaining the actual reason behind the constraint works better than the hard rule.
- HTML artifacts have become Anthropic's default way to share plans, PRs, and status reports internally — not just documents, they're a living reference the agent can also read back later.
- A cheap HTML prototype of a design proves out the concept before paying for the expensive version — in this case, HTML mockups before rebuilding the same overlays in React.
- Multiple Claude sessions in Slack (Claude Tag) each keep their own memory per channel, letting one person run background PR babysitting, spec exploration, and initial investigation in parallel while a single active Claude Code session handles the one task they're focused on.
- The habit that saves the most time isn't a tool — it's picking one task to actually focus on, because the time lost to a lazy, half-supervised prompt tends to cost more than doing it right the first time.
- Becoming more technical in an AI-driven workflow isn't about learning syntax — it's about learning the tradeoffs of the systems underneath (encoding formats, transcription libraries, backend services) so you know what's actually possible.
- Whisper transcription has known, specific failure modes — silence can turn into hallucinated phrases like 'thanks for watching,' words can split across chunks, and it has no built-in speaker recognition — and knowing them up front prevents building a workflow around assumptions that don't hold.
Three habits that get Claude to work longer without babysitting.
WHAT TO LEARN
Locking in an exit condition, splitting the agent that does work from the agent that grades it, and treating planning as an ongoing search for unknowns are what let one person run Claude far past a single prompt.
01Goals, loops, and workflows to make Claude work longer
- /goal locks in an agent's exit condition so it won't stop early on a task you've already fully scoped — it's built for handing off execution, not for exploration.
- Workflows are worth the overhead specifically when 'good' isn't a deterministic check — give the agent a Figma file to verify against when you can, and reserve a rubric-plus-verifier workflow for fuzzier judgment calls like a screenshot match.
02Live demo: Creating a polished video with one prompt
- One-shot demo: a single prompt (record a video, transcribe with Whisper, build captions and overlays with Remotion, don't stop until it's rendered) produced a finished, captioned video end to end.
03Plan by removing unknowns before you build
- Planning with an agent is not a one-time handoff — it's an iterative loop of exploring, building something small and cheap, and re-specifying once implementation surfaces what nobody anticipated.
- The real goal of a planning pass is finding your unknown unknowns — research into how a system works (like Whisper's specific failure modes) counts as planning even when it produces no spec document.
04How Anthropic uses HTML artifacts for planning
- HTML artifacts have become Anthropic's default internal format for sharing plans, PRs, and status reports — cheap enough to prototype in, and readable enough that people actually read them instead of glazing over a long markdown file.
- Prototype cheap before you build expensive — an HTML mockup of a video overlay design proves the concept before paying to rebuild the same thing in React.
05Running parallel Claude tasks through Slack
- Claude Tag (Claude in Slack) keeps separate memory per channel, which lets background investigation, PR babysitting, and specs run in parallel while one focused Claude Code session handles the single task you're actually watching.
06Using separate creator and verifier sub-agents
- A model grading its own output is measurably more lenient than a separate agent grading it — split worker and verifier into different sub-agent sessions whenever quality is subjective.
07How Thariq stays focused while Claude works in parallel
- The habit that saves the most time isn't a tool — it's picking one project to actually focus on, since a lazy, half-supervised prompt tends to cost more time later than doing it right up front.
08Why Anthropic cut Claude Code's system prompt by 80%
- Anthropic cut Claude Code's system prompt by 80% because smarter models need fewer examples and fewer hard constraints — heavy example lists and blanket 'never' rules end up over-constraining behavior.
- A blanket 'never do X' instruction is usually a lie, since you rarely mean never — stating the actual reason behind a constraint tends to work better than the hard rule itself.
09How to become more technical when AI writes the code
- Becoming more technical in an AI-driven workflow isn't about syntax — it's about learning the tradeoffs of the systems underneath (encoding formats, transcription libraries, backend services) well enough to know what's actually possible, and it takes deliberately pushing the agent to explain itself.
Terms worth knowing.
- /goal
- A Claude Code command that locks in an explicit exit condition for a task, preventing the agent from stopping early once the user has done enough scoping that they just want it executed.
- Workflow (Claude Code)
- An orchestration pattern where a main agent splits a task across sub-agents that do the work in parallel, plus a separate sub-agent that verifies the output against a rubric.
- Self-preferential bias
- The tendency of a model to be more lenient when judging its own output than when a separate model judges the same output — the reason worker and verifier roles are split across different agent sessions.
- Claude Tag
- Anthropic's Slack integration for tagging Claude directly in channels or DMs; each channel or thread keeps its own persistent memory, letting a user run several background Claude sessions in parallel.
- Multiclouding
- Running several Claude sessions on different tasks at once — background investigation, PR babysitting, and specs in Claude Tag, with one active focused session in Claude Code.
- Whisper
- An open automatic speech recognition model used here for transcription; it predicts the most likely text for a stretch of audio rather than transcribing with certainty, which produces specific, known failure modes.
- Remotion
- A React-based framework for building and rendering video programmatically, used here to generate word-highlighted captions and overlays from a transcript.
- HTML artifact
- A self-contained HTML page Claude generates to represent a plan, design exploration, PR summary, or status report — used internally at Anthropic as a shareable, re-referenceable format.
Things they pointed at.
07:50toolWhisper
05:00toolRemotion
02:10toolFigma MCP
31:40toolSimplify (Boris's repo-cleanup tool)
Lines you could clip.
00:01
“As the models have gotten smarter, they need less direction, fewer constraints, and fewer examples.”
01:35
“Slash goal is one of those things that it helps the agent remind itself what its exit condition is, and only allows it to exit once you know it.”
14:20
“The prompt box can definitely just be a lazy button, where you're just like, hey, just do the thing. But usually you end up paying for that.”
28:30
“We call it self preferential bias. When a model prefers its own outputs, it's going to be more lenient at verifying it.”
34:10
“We cut down the Claude Code system prompt by 80%. The reason is that as the models have gotten smarter, they need less direction.”
37:50
“The goal of learning to be more technical is to know my unknown unknowns.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00As the models have gotten smarter, they need less direction, fewer constraints, and fewer examples. We've got slash loop.
00:07We've got slash goal, and we've got workflows. These are all geared at trying to get the agent to run for long periods of time. This was just one shot, like, when a one prompt.
00:16And so it's created the caption, and it's created the little overlay, and now it's going to fade to black. My goal for this year is to be more productive but work less.
00:27Hey, Ron. Today, I'm really excited to welcome Tharek from the Cloud Code team. This has been a long time coming, so I'm gonna ask Tharek to show us how to design loops and workflows to get Cloud to work longer and how he personally uses Cloud Code and also the new Cloud Tech.
00:40Welcome, sir. Hey, Peter. Yeah.
00:42Thanks for having me. I've been been excited to to join for a while. Yeah.
00:45Alright. So may maybe before you demo anything, maybe you can talk about, at a high level, how you and the CloudCall team think about how to go from you being the person prompting the agents to kind of design loops and some of the stuff to get the agent to kind of work by itself.
00:58Like, how do you guys think about this? Totally. Yeah.
01:00Yeah. So I I think loops are a pretty general
01:04term, I I think, for just different ways of, like, having the agent get feedback or, like, sort of, uh, yeah, work for long periods of time in some in an orchestrated way. And so we've got slash loop, we've got slash goal, and we've got workflows.
01:18And these are all, yeah, like you said, geared at trying to get the agent to run for long periods of time. Slash goal is one of those things that it helps the agent remind itself what its exit condition is, right, and only allows it to exit once you know it.
01:33And so goal is really great for when you, uh, have a complicated task and you really need to, you know, make sure that this is, like, done at the end. You really want to prevent any stopping.
01:45You kind of also want to give the agent feedback a signal to basically say, like, hey. Power through things. Right?
01:51And and so I I think one of the like, if you were to step into Claude's shoes sometimes, you ask someone ask you to do a task, you run into a like, maybe, you know, sort of a complication or something that is maybe not quite up to spec or, like, it doesn't match what the user said. You might stop early and be like, hey.
02:07Should I keep doing this or not? Right? And so goal is one way of just kind of the user indicating like, hey.
02:14I've I've done enough spec and exploration. Right? I understand the problem space.
02:19Uh, just go execute on it. And if you run into something, you know, fill it in. Right?
02:24So that's how we think about goal. And and then workflows are, in my opinion, maybe, like, the most powerful form of this where you can, you know, spin up sub agents to both do the work, paralyze the work, and verify the work. And and so I think especially for nontechnical work, this can be, like, a really great way of taking a nondeterministic task and then breaking it down roughly into a deterministic task.
02:46So, yeah, I I think that's how we think about some of these. And, you know, I I think I'm gonna do a quick demo, but, like, before that Yeah. Like, just on the Clocko team,
02:53how do you guys use some of this stuff? Like, do you do, like, a go make this match design or, like, go make this
03:00number go up? Or what do you use? Yeah.
03:02Yeah. I mean, everyone does different things, right, based on, like, kind of the, like based on, yeah, what what they're trying to do.
03:11I I think, like, you know, Jared's talked a lot about the the bond rewrite in Rust and how that use workflows, and and he's gonna be talking more about that.
03:20I think, like, anytime you have a deterministic sort of, like, signal, like, for example, like, latency. Right?
03:27Like, goal is, like, a great thing to just sort of, like, have the agent explore things sort of in, like, an auto research way. Right?
03:36And, yeah, I I think that, like, you like, let's say you're talking about design.
03:43Right? I think one way you wanna think about this is, like, how well can the agent, like, understand what your, like, design is to begin with.
03:52Right? And and so you might want to, like, turn it into like, if it has a spec that it can verify.
04:00Like, if it's a Figma file, you use the Figma MCP, and then you're, like, slash goal, make sure that the rendered design matches the Figma MCP. That's a lot easier than, like, a screenshot. Right?
04:11And and so if you're a screenshot, maybe you want to do a workflow where you like, it's more squishy.
04:17You know? Like, there's, like, a rubric that you're evaluating against, and you have a verification agent and and things like that. So Got it.
04:24Yeah. Very dependent on the task. So, basically, the planning matters a lot, right, to give it because I tried giving it a goal to, like, just build me an amazing game, and
04:33and then it kinda went off the rails because, like, you know, it's just, one one line. So Yeah. I mean, I think the
04:39the details of a a lot like, you know, there's a lot to figure out in what you want.
04:45You know? There's a lot of work that goes into that. Yeah.
04:48Got it. Okay. So let's make this really concrete.
04:51You've been sharing really awesome videos on Twitter about, like, using Cloud to make the video, And I think you have, like, a video work workflow that you wanna share with us. Right? Yeah.
05:02Exactly. So I yeah. I've had a few people, uh, ask about this, and I think this is a good sort of demo both for, like, video editing, but also just, like, how to think about, like, a nontechnical work.
05:12So ten minutes before I got on this call, I recorded a quick video. And so this is me just, like, you know, making a video. Basically, what I'm saying is, hey.
05:21It's me. You know, like, I'm on the Peter Yang podcast. I'm going to point to where I want an overlay to show up, and then I'm basically saying, like, hey.
05:32Fade to black. You know? So this is, like, a quick video that I that I put together.
05:36Right? And now, like, the prompt I gave it is this.
05:42Right? Like, this is a repo for Peter Yang podcast.
05:47There's a sample video here called Peter Yang recording, transcribe it using Whisper, then use Remotion to create a UI that shows the transcript with each individual word being highlighted and different overlays. And, yeah, goal, don't stop until the video is fully rendered. Right?
06:02Nice. And so now I like this was just one shot, like, with a one prompt.
06:08Right? And what it gave me is this. So it's it's transcribed it.
06:12It obviously doesn't know my name is Tarek. Right? And it's, like, overlaid the the the text, or sorry.
06:20It's created the caption, and it's created a little overlay, and now it's going to fade to black. Right? And so Oh, awesome.
06:26Okay. So so you basically sent out the instructions out loud. Right?
06:29I sent it out loud. Yeah. Exactly.
06:30And and, of course, there are many ways that you can do this, but it's like now in this directory, you know, uh, transcribed everything. Right?
06:38Yeah. Uh, using Whisper. Uh, it's created the Remotion stuff.
06:41Yeah. This is, like, the very basic start to, uh, video editing workflow that I have, which is, yeah, I think a good signal for, like, how to do nontechnical work in, uh, in Cloud Code.
06:52And so you kind of have to dismally with the prompt stuff, but then you probably save it into a skill or something. Right? Yeah.
06:57So right now, I don't have a skill. I I think one of the things I try and do first is really figure out what I want before I turn it into a skill. You know?
07:05And and so what are the, like, edge cases here? Um, one of the things I noticed that I had to prompt it towards is is getting better at figuring out where my, um, where my hand was.
07:19Right? And so, like or where I was pointing towards. And even now, I'm not not thrilled with, like, where this showed up.
07:27You know? And so, like, yeah, like, you could imagine that one direction I could take this in is I could be like, oh, I want to track my fingers or track my face and then get give the agent more metadata, right, so that it can then do interesting overlays.
07:41This episode is brought to you by WhisperFlow. WhisperFlow saves me at least three hours a week and is one of my favorite AI apps by far. It's just so much faster to dictate to AI using your voice than to type.
07:53You just talk naturally, and it outputs clean, ready to send text. WhisperFlow even removes filler words and formats your sentences for you. I use WhisperFlow for everything, including drafting newsletter posts, writing product specs, replying on Slack, and more.
08:06It works on Mac, Windows, iPhone, and Android across all of your favorite apps. Try it free at whisperflow.com, and use my code peter whisper flow to get six months free.
08:17That's peter whisper flow. Now back to our episode. I think you had another tab that had, like, a plan or or something.
08:23Is that for the video scale? Yeah. Yeah.
08:25So I I think I want to talk I'd love to talk a little bit about how I got here. You know what I mean? So, like, I think, like, even to do this
08:33prompt, I think it took a bunch of planning and learning to get here. Right?
08:39And then I think that, like, one of the things that I I I think when we talk about plans, we often are talking about things where it's like just one shot.
08:51Right? Like, you plan and then you do something and that that's it.
08:56But I think planning is more of this iterative process of, like, exploring, investigating, finding out what you don't know, what you want.
09:05Right? And then, like, together, like, in the end, that sort of simplifies itself. But so, like, here's an example.
09:13But while we do this, I I would love to do something where it's like, okay. I'm going to give it okay.
09:19I wanna say, like, I want to update the UI of these overlays. Right?
09:26And I want to use I want to use Peter Yang's style.
09:35Here's his blog. Create a HTML artifact for exploring different designs for the overlays and captions, design variations.
09:52Okay. So this is, like, an example. To me, this is also planning.
09:56Right? Because, like, what I'm doing is sort of, like, I I'm trying to figure out what I want here.
10:02Right? And and so it's more of, like, exploration. You can see one of the important things I'm doing is, like, I'm giving it a reference.
10:08Right? So when I give it, you know, your website, it can now fit fetch the HTML and start doing this.
10:15So, like, this is sort of, like, a plan that will will run while I sort of go over the other, you know, kind of things I do when I'm planning. Right? So one of them is like, okay.
10:23What how does transcription work? Right? And and this is actually, I think, pretty important to know because, like, the the it informs a lot of the edge cases.
10:31So this this transcription uses Whisper. And so, you know, like, Cloud Code sort of put this explainer together, which is honestly, like, kind of amazing.
10:42Yeah. The important thing for me is, like, the like, ways things can go wrong.
10:48Right? And this is something I prompted it in the plan. It was like, want to explain Whisper to me and understand what the edge cases are.
10:54Right? And so one of the things that says, like, silence can become thanks for watching. Right?
10:59Or, yeah, a word can be split into two chunks. It doesn't know it doesn't have speaker recognition. Right?
11:05And so there are a bunch of different, like, edge cases here where I I think all of this is quite, like, good, and this sort of, like, helped me build a confidence in using Whisper. Right?
11:17Yeah. Um, but having these edge cases and knowing sort of what the limits are upfront really helped me, like, avoid this case where I I build this complicated workflow during Whisper, and then, you know, I realized there are things going wrong.
11:31Um, and I, like, sort of didn't have these unknown unknowns. Right? And so this case is sort of like me sort of doing planning, but, really, it's like me discovering my unknowns.
11:41And I think that can take a lot of different shapes. It can be learning.
11:46It can be technical specs. Right? It can be mock ups and exploration.
11:51But, yeah, I I'm starting to, like I feel like the word plan is maybe, uh, too broad right now. So It's it's it's more like exploration and understanding, like, you're trying to do, basically. Yeah.
12:02I like to say, uh, like, getting rid of your unknowns. You know? Like, I think whenever you have a task, it it kind of, like, almost always, uh, there's a lot that you don't know.
12:14Either you don't know how things work or what you want, and it's very, very iterative. You know? And so it's not just like a you write it all down once and then you implement it.
12:23I think there's, like, many steps and different passes. So And Yeah. And and and, dude, this is, like, a beautiful plan, man.
12:29Like, do do you have like, is this, like, the cloud template or something? Like, it has images? Using the front end design plugin, I think.
12:35You know? Like, I think, like, uh, yeah. It's not, uh, yeah.
12:39There are uglier plans here. Like, I've gotten this, like, remotion plan. Yeah.
12:44I I think that, like, know, the the design of it doesn't matter so much as sort of, like, you really just want to make sure it's something that you really do read, you know, and kind of get a sense of, like not all of it, but, like, there's important parts in it. And I think one failure motto see is that people still sort of, like, glaze over, you know, the plans and explainers.
13:04So, um, yeah. Yeah. It becomes really because, like, AI can write all these, you know, crazy markdown files, and they're usually pretty long.
13:10And then Yeah. At some point, like, I just get lazy. I'm like, okay.
13:12Well, you know, just just do it. Just do it. Yeah.
13:15Yeah. Yeah. Exactly.
13:15I I think that's, like, the thing that you know, it happens to all of us. Right? Like, the a like, , right, where you're just like, hey.
13:23Just do the thing. Yeah. But, usually, you know, like, you end up paying for that.
13:30Right? Because if you're if you're trying to do something serious and, you know, like, you're taking the lazy step at each way, it'll it'll it'll end up taking longer, maybe costing longer too.
13:40So
13:42So it's kind of this iterative process. You're asking about, like, the the whisper, and then you kind of, like, learn a little bit more, and then you ask it to make another plan.
13:49And and at some point, you have a good plan that you can share. Right? Yeah.
13:52Exactly. Yeah. Yeah.
13:53So I I I I did a lot of for this video recognition stuff, I did a lot of, like, you know, research into how different video, like, algorithms worked. Right?
14:03Like, I I think I did one on, like, video segmentation. I wanted to put, like, text behind the subject. And so this was, like, something I, like, explored and learned more about.
14:12And I ended up finding that, like, there wasn't something reliable enough here for me to use. But, yeah, I I think there's, like, just a lot of this, like, iterative process of, like, finding out what do I want, you know, and what is possible, right, before I can do it.
14:26And so, um, it looks like this overlay is maybe almost done, so we'll we'll we'll
14:30see it soon. Yeah. And how about, like, on the team?
14:33So have have you got the rest of the team to
14:36review HTML? Is that a markdown? Or Yeah.
14:38Yeah. For sure. I mean, I I think, like, everyone is different.
14:41Right? So I think that, like but I think HTML artifacts mean, we launched artifacts. Right?
14:46Currently, it's only on Teams and Enterprise, but hopefully coming to Max and Pro. And, yeah, that that's, like, how we share things now at the company.
14:57It's like we ask Claude to make an artifact, and this can be, you know, of a plan, of a PR that we've already done. It can be, you know, like, a status report, like, an incident report, things like that.
15:08So, yeah, it's definitely how we do it.
15:12Yeah. Alright. Let's see what I generated here.
15:15Yeah. Let's let's do it. So let's say open overlay style HTML.
15:20Okay. So, yeah, here are some of the, like, you know, the options that's made.
15:25I think it's probably maybe, like, indexed more on the stack over or or the, uh, Substack brand versus your brand. You know what I mean? Because I don't maybe I don't have a brand.
15:34Yeah. Yeah. Yeah.
15:35Yeah. No. I think you do.
15:36Right? I think you do. Like, I I think you've got the red and white.
15:39We can probably iterate on it there. But this is one of the things where I'm like, you know, there there's quite a lot of difference between these different, like, you know, designs.
15:48Right? And Yeah.
15:50Yeah. One of the ways I like to to plan is is to do this exploration, especially because I'm not a designer. So, like, I just really just only know it when I see it.
16:00You know? Got it. Got it.
16:02Like, just to wrap up the plan conversation, like Yeah. You know, I've been writing, like, product specs for, like, ten years, and, like, you know, they usually have, like, what's problem we're solving, what's the solution, what's the goal, so on and so forth. Right?
16:12But I feel like part of this stuff is, like, read by agents now. And I I I feel like the different sections of your spec needs to change, or or or maybe, like, the prospect and technical spec is, like, one thing. Like, how how do you guys do that on the team?
16:24Like, you have, like, one session for the humans to read and one session for the AE tools to read, or or how do you think about it? Yeah. Good question.
16:30Good question. I I think that it is very tied together. Right?
16:33Like, I I think a spec kind of can evolve even towards,
16:37like, planning. Like, I I think one of the things I think I'm showing here let me see.
16:44Implementation. Yeah. I I think, like, one of the things that you sort of see here sometimes is that when you run something, the the model can find things that it didn't anticipate or you didn't anticipate when you're implementing.
17:00Right? So I see. This is like, I I don't think of specking just happening at the start.
17:06Right? It's like you you start with, like, the human request, and then the agent does some, like, technical exploration.
17:13It comes back. Maybe you do some mock ups, some explainers trying to understand your unknowns. You refine that.
17:19You give it to the agent again. It might start implementing it. I ask it to keep implementation notes as it goes so that it, like, finds out, you know, like, what are the you know, what are things that we were not expecting about this implementation.
17:33And once we have that, we can actually, you know, respec if we need to. Right? Like, depending on how things go.
17:39So, yeah, I I think it's much less like there's, like, one handoff of spec to implementation and more of like, back and forth process.
17:47Yeah. Yeah. Because because I guess it's pretty cheap to build now, so you can just, like, maybe ask it to build a simplest version of this and, like, there's probably a lot of bugs and issues, and they're keep iterating.
17:56Right? Yeah. Exactly.
17:57Or, like, the prototype version of this. Right? Like, that's what how I think about these overlays.
18:01These are, like, prototypes of the, you know, of the design. Right? And Okay.
18:06If we, like like it, then we can do the more expensive version, which is instead of HTML, we can do it in React. That means you have to, like, you know, sort of rerender the video and do all of the code changes.
18:18So, like, what's the, like, smallest
18:21step you can take to prove out the concept that you want to, like you prove out the spec more. Yeah. And, like, I yeah.
18:28I love how it takes implementation also. So the next time you start a new section, it can just refer to the HTML. Right?
18:33. Yeah. Yeah.
18:35Exactly. Yeah. Yeah.
18:37Got it. What percent of your work now is done through Cloud Tech on Slack versus in, like, the terminal or, like, you know, the Cloud app?
18:45Yeah. I I think that, like, the way for me it's trending is that, like, there's a lot of parallel stuff happening in CloudTag. And so any multiclotting is often happening in CloudTag unless there's a reason why it needs to happen locally.
18:58You know? Um, I think one of the things we've put a lot of time into is, like, making sure that our environment can run, um, in in remote.
19:05And so, um, yeah, I I multiclouding is happening in CloudTag.
19:11You know, initial expirations, PR, specs, like, trying to understand something is happening in CloudTag. And then once I get into there's usually one thing I'm focusing on, and that's happening in Cloud Code.
19:22And it's, like, more, you know, like, back and forth iterative. Right? Okay.
19:27And so, yeah, I think I think that's how I'm using it by I think everyone at the company has sort of, like, different, like, approaches and mixes to it. And when you say, like, what do you mean by multiclouding?
19:38Oh, yeah. Like, just like if I have, like, multiple tasks happening at once, I'll especially if I'm, like, trying to involve someone else. Like, if I am, um, you know, I've got a PR that I'm trying to get merged, you know, like, what one thing I'll do is, like, I'll, you know, ask it to babysit, uh, the PR fixed test and then tag a reviewer, and my, like, reviewer will get tagged in the same Slack channel and will, like, be able to talk there.
20:03But just generally, multiclouding is, like, anything any background work where before I was having, you know, like, five different clouds, uh, in my cloud code. Now, like, it's mostly one active cloud Code session and then a bunch of Cloud Tag sessions.
20:15Oh, interesting. Okay. And, um, so you talked to Cloud Tag both through, like, DMs on Slack and also, like, in shared team channels, I'm guessing, or
20:25property. Yeah. Yeah.
20:26In in Slack. Yeah. Yeah.
20:28Both in my private like like, have a yeah. My Tharek Claude Slack channel where I do, like, most of my work. And then, uh, yeah, we have, like, team channels, like feedback and things like that where we, um, or, like, project specific engineering channels as well.
20:42That's, like, a common pattern we see. Yeah. I guess it makes sense because I I do think, like, these coding apps right now are, like, primary single player experiences.
20:50Like, you're kind of talking to the agents through different threads. Yeah. And and if you think about it, like, the the cloud app actually kind of looks similar to Slack.
20:56There's, a bunch of threads. Like, it's just like Slack has a bunch of channels. So I guess, like, the idea is, like, Slack is now the multiplayer cloud experience.
21:03Right? Because, like, everyone's there already.
21:05Yeah. I mean, I I think, like, this is, you know, where it's starting. Right?
21:08Like, I I think, ultimately, we think cloud will be, you know, this, like, sort of proactive agent that's sort of, like, meeting you where you are. You know? And I think Slack is where Anthropic is.
21:16Right? And, yeah, it's a very natural way to explore. Yeah, it's surprisingly good at coding.
21:22It can, like
21:23you know, there are some people who just do all of their coding almost completely in CloudTag. Yeah. But okay.
21:28Maybe this is a dumb question. But how do you, like you know, in CloudTag, like, easily triggers scales and stuff? Like, if I'm in CloudTag, like, do I have just tag Cloud and slash the scale, or how is it?
21:38Yeah. You can just tell it to use the skill. Okay.
21:41Okay. That makes sense. Yeah.
21:42Yeah. Yeah. There is still I I think there's still UX iteration we're working on around all of this stuff.
21:48Yeah. Okay. I guess, do you think, uh, do you agree?
21:50Like, I feel like the future for this, like because you're you gave me a talk, right, on on human agent interaction. And, like, um, I feel like the future is just like the agent is just like another employee. Like, you got a you got a onboarding employee.
22:02You can talk to it through Slack. You can, like, give give it a phone phone call.
22:06You know? What do you think? Yeah.
22:08Um, I think it's a good question. I I think that, um, I think these metaphors can sometimes be helpful but also limiting in some ways.
22:16I I think that, like, one of the things, you know, when we talk about identity is that, like, in CloudTag, every agent every channel has its own memory. Right?
22:26And so that is, like, I think, one choice. Right?
22:30Like, you could also imagine that there are, like, you know, multiple clouds you tag. Like, each one has a different Slack identity, and you're tagging them and things like that.
22:39And and I'm not like, I I think it is kind of helpful to think of an agent that is, like, sort of has persistence and, you know, uh, memory, but, um, I think they're also different from coworkers in in in some ways.
22:55Right? But I think they're, like, they're proactive and, like, you know, the, yeah, it it it's proactive. It has memory.
23:01It has identity. Um, but, yeah, it's, like, uh, just sort of like an evolution of Cloud Code.
23:07And I think we want to see where, like, you know, uh, where the models will
23:13will take us there versus sort of, like, putting it into a box. You know? Yeah.
23:16That's a point. Yeah. Like, it definitely has a lot better memory than any human.
23:21Better and worse sometimes. You know what I mean? Like, it's spiky.
23:24Right? So yeah. Yeah.
23:26It makes sense. Okay.
23:27Alright, dude. Well, why don't I do something fun? I I wanna show you my Cloud Code setup.
23:30I think we talk about this in person. I feel pretty proud of this. I I built this podcast production scale.
23:35Mhmm. And, uh, what it does is, like, it takes a transcript of an interview that I did. I I actually interviewed your colleague, Jess, and, uh, it it kind of generates a bunch of stuff for for me.
23:45So it generates thumbnails and, like, stuff to cut. So, basically, I I just tag it and I paste the transcript, and then it starts generating, like, you know, clickbaity YouTube thumbnails and stuff.
23:56Yeah. Yeah. So it generates the stuff.
23:58And and, like, I I basically give it, like, my examples to try to keep it on track. Yeah.
24:04But I guess that this skill is kind of, like, trying to do a lot of things. So maybe, like, I'm curious if you have any feedback or you know? Yeah.
24:09This is cool. Can I see what it produces? Like, the the clips or, like, you know Oh, what it produces?
24:14Yeah. Yeah. So some of it what produces is is just, like, pure text.
24:17It's like, here's a newsletter post. Here's some takeaways. Copy into YouTube.
24:20Yep. Yeah. Then I have this other scale called the video postcodes.
24:23It's kind of similar to yours, but I'm I'm sure it's not as sophisticated. But, like, it basically takes the YouTube video for this thing Cool. And then it, uh, extracts it, and and then it gives me some ideas for clips to to make.
24:36Yeah. Yeah. Does it do the clips as well?
24:38Yes. It actually does. So I I said do two, and then it uses I don't actually know what it uses.
24:43It a bunch of random stuff too too. Probably FMFMPEG and stuff. Right?
24:47Yeah. Yeah. Probably FMFMPEG, and then it makes a it makes a video.
24:50Yeah. Oh. And it has captions.
24:51Yeah. Yeah. That's awesome.
24:53So so yeah. This is not perfect, though. But but but yeah.
24:56Mhmm.
24:56Yeah. I mean, um, this is great.
24:59How how, like, how well do you feel like it matches what you want?
25:07You know? Like or do you know, like, what it would look like to be better?
25:10I wish it could, like, ask them, like, you know, b rolls and the stuff that you showed me, like, overlays and stuff like that. And, like, I I I wish it was smart enough to know how to pulling, like, for for the video clip, like, you know, if I want the Cloud logo or or something, you can just find it online.
25:26Mhmm. Yeah. Yeah.
25:27So should I should I just tell Cloud Code to do all this stuff? Or I I think you can. Yeah.
25:31I mean, I think that, like are are you working in one persistent repo? Or, like, I think one of the things I think about is, like, sometimes you want a skill.
25:39Sometimes you want a repo with a lot of scripts and things that you are, like you know, like, it's more of, like, a a workspace that you're working in. Right? And, like, I think of a skill yeah.
25:49Like, I I think sometimes a skill can be instructions on how to create that workspace. Right? And so Yeah.
25:55That that's maybe something to think about because, like, the more, you know, scripts and sort of, like, things that you build up, the more the agent can the less it has to do from scratch. You know?
26:05But yeah. I mean, I I have my folder each time.
26:09No. I have all my skills at the user level. Okay.
26:13Yeah. Yeah. And and and there's, this personal or whatever personal OS folder that has, like, all the output and stuff.
26:18I see. I see. Okay.
26:19Cool. Yeah. Yeah.
26:20Yeah. But but but you're saying, like Yeah. I mean, I think you can sort of build, like, a video editing harness here.
26:24I I'm I'm
26:26curious about the thumbnails. Do you ever use, like, an ImageGen API, like a Gemini or even an OpenAI or something like that? Yeah.
26:34I found it really bad at changing my face. So, like, for example, if I have a smiling thing and then ask to make a, like, a shocked face, it makes me look so broad, dude.
26:43I mean
26:45That's funny. That's funny. But Yeah.
26:48Yeah. Yeah. But I think it's pretty good at, like like, I don't you know, if I give up my face and it tells you to change the background of the text, it's pretty pretty good.
26:54Yes. Yeah. Yeah.
26:55I mean, I I think Claude is quite good at like like, one of the things that Claude is good at is using other tools. So, like, you could give it, the Gemini or OpenAI image gen APIs. Yeah.
27:05And then you can ask it to look at your faces that it generates as well, you know, and sort of, like, tweak it there. So it can do this, like,
27:14interactively and per progressively as well, which which can be helpful. You know? Oh, okay.
27:18Interesting. Yeah. So so, basically, basically, what I do is I I basically just chain a bunch of skills together.
27:23Like like, know Cool. Prepare this thing and then use a thumbnail and so on and so forth. I feel like I I know how to use skills now, but, like, some of the other stuff that you talked about online, like dynamic workflows and stuff,
27:32like, have no clue. Right? I have no clue.
27:34Yeah. How would we use a workflow for for something like this? Right?
27:37I I think, like, one thing you can imagine is, like, when you're using I I think your thumbnails example was a good one where or sorry. Your shorts example was a good one, where maybe you want to generate 10 different shorts or five different shorts or something.
27:53Workflows are a great example for that. Right? So you can, like, say, hey.
27:57Like, um, the main agent decides which five areas of the clip, you know, you want to, like, create a short for. Right?
28:06And then the, like, workflow spins off a sub agent for each one. You also give it maybe a rubric for, like, what does a good clip look like.
28:14Right? Like and then each one is verifying against it, And then you're getting back at the end this, like like, you know, each clip has a maximum amount of compute put into it to make sure that it's matching what you want versus, like, if you're doing two or three simultaneously sometimes, like, I I think, like like, Claude might, like, verify or put less work into any individual clip.
28:40So I see. And and to create a workflow, do I just, like, just tell to make a workflow, basically?
28:46Yeah. Or Just Just be like, hey. Like, yeah, you know, create 10 clips here, use a workflow.
28:52This is my, like, rubric to to verify, like, what a good clip looks like. Right?
28:59And this can also be a skill. So you can package a skill with the workflow, and then, like, you can the workflow is just a JS file.
29:07And you can, like, ask it to, like you save the JS file into the skill, and now you have this, like, reusable skill that you can use. And this is, a dumb question, but, like, the main advantage over doing this workflow thing versus just using a scale is, like, it can spin up sub agents and, like, keep the context clean? Or Yeah.
29:25Yeah. Yeah. So I I think it's like yeah.
29:27One of it is context. I think the other one is sort of, like, laziness and verification.
29:34Right? So, like, if you have a you know, like, a a short is one example of something where you don't have a deterministic, like, hey.
29:44Is this a good short or not? Right? And so if you have a rubric and you have a, like, you know, verification agent that reads the rubric and then it looks like, oh, like, hey.
29:56Let me review the short to make sure it's good and give feedback. Right? That's something that the workflow enables as well.
30:03And and we found that, like you know, we call it self preferential bias. Like, when a model prefers its own outputs, it's it's going to be, like, more lenient at, like, uh, it's going to be more lenient at, like, you know, verifying it.
30:17So Okay. So, basically, you wanna have a separate cloud doing the work versus verifying the work. Right?
30:23Yep. And in this case, like, a separate cloud coordinating the work, uh, so that your main agent is coordinating it coordinating it, each one doing the work and and then verifying.
30:33Yeah. Okay. And that's why, like, the three separate clouds have, like, different context windows, and it's not just, like, less biased.
30:38Yeah. Exactly. And and they will both all use more compute, I I think.
30:41Like, they'll put more, like you know, less likely for it to, like, stop early. Yeah. Think more.
30:46Yeah. Okay. Got it.
30:49Oh, awesome, dude. Well, let me ask you some, like, high level questions, dude, since since, uh, it it looks like you do you mostly use Clot in your in the terminal, or do you actually use the app?
30:57I'm a mix of terminal and desktop. Yeah. Depending obviously, we do a lot of, like, what we call ant footing.
31:03So it's sort of like, you know, whatever I feel like I need to test the most. Yeah.
31:08Got it. Do you have any tips for
31:11like, I I I feel like sometimes I get exhausted, man. Like like, there's, like, five threads going on at once for five different things, and then, like, I'm I'm just like, they're constantly paying me.
31:19It's actually worse than having back to back meetings in some way. Have any tips for how to keep your own windows clean.
31:26Or you know? Yeah. I mean, this is a good question.
31:29You know? I I think that this is like my goal for this year is to be more productive but work less. You know?
31:36And so I think this is, like I think something we should all sort of be able to push ourselves on. I think that, like, you know, like, I I think one thing we try and do is sort of, like I try to have, like, one project I'm focusing on.
31:52You know? And and so even if there are other things that I just need to get unblocked, like, I need to have this build and merge and, you know, things like and explore something. Like, I think one project that I'm really focusing on is very helpful just, uh, because I find that the thing that cost me the most time is when I'm like, yeah, maybe a little bit lazy.
32:11And and now I've, like, done a bunch of I'm, like, sort of multitasking these things. I do a lazy prompt, and then I'm like, oh, like, I've wasted this time now.
32:19You know? And then so, yeah, I I think that there's, just because like, I think there's a optimal amount of multi multi clouding depending on, you know, who you are, what you're doing.
32:31But for me, at least, it's something where, like, I tend to have, like, one task where I'm really focusing on the most. Yeah. Okay.
32:36So maybe, like, the AA agents take care of the rest or something. But, like, one last time. Or even prioritization for yourself.
32:43You know what I mean? I I think can be, like, a good, like, angle. Yeah.
32:46Makes sense. And and and, like, do you have concerns? Like, I I worry that,
32:51like you know, when because the agents, like, work pretty hard and they keep going. And and, like Yeah. May maybe this is my problem.
32:57Like, I I don't actually like like I said, I don't actually read everything that it preproduces carefully. Oh, sure. So that's then I worry that, like, the whole report just turned to slop at some point.
33:04Like, if I don't Mhmm. Like, do you have any, like, routines or jobs that clean this stuff up periodically?
33:10Or
33:11Yeah. So, I mean, obviously, we have things like simplify, which Boris put out that, like, you know, sort of simplifies the repo.
33:19I think it depends on what you're using it for as well. You know? Like, I I think someone like when you're using it for the outputs, like, you're using it for the, like, the video outputs, um, those are pretty, uh, like you know, maybe the quality of the code matters less.
33:38You know? And the agents are very, like, you know, persistent, and and they'll figure it out. But, um, I find organizations often more for me than the agent to sort of, like, you know, make me feel better about the workspace, um, if I'm only caring about the outputs.
33:51But, yeah, you can always just ask it to to simplify or organize.
33:55Um, Yeah. Okay. Okay.
33:57Got it. And and you try to keep this con do do you try to, like, minimize, uh, and try to keep the concept window clean? Like, you try not to, like, have a bunch of, you know, MCPs on or, like, you know, a cloud that empty has a super long like, do you try to optimize that?
34:10Or
34:11Yeah. I mean, I I think that, like, yeah, I yes. I I think that one of the things we've noticed, especially as the models have gotten smarter, is that, like you know, we'll be talking more about this.
34:23We we cut down the Cloud Code system prompt by 80%. And the reason is that as the models have gotten smarter, they need less direction.
34:33They need fewer constraints, and they need fewer examples.
34:37You know? And so a lot of our system prompt would be like, okay. Like, here's the Batchel.
34:41Here are, like, five examples of using the Batchel. Never do it in these cases. Right?
34:46And the models are now kind of aligned enough that they know, like, hey. We don't need to, like we we like, we don't the examples are almost constraining it, you know, because now it's like, oh, like, you want things like this example.
35:04And so if if you remove the examples, it actually can be more free format. And, like, the constraints can also constrain it too because, like, you oftentimes, when you say never, you don't really mean never.
35:16You just mean, like, most of the time, don't do this. And if you give it the reason you don't want to do it instead, that can be more effective than the, like, don't do this constraint.
35:26So, um, I think that, like, all this to say is that, like, you want to trim your context. I feel like Cloud.
35:32MBs are probably too long right now, and you probably want to shorten them more and more. Right? Uh, so probably a lot of skills are too long.
35:41Um, I think MCP is you know, depends on the MCP. I I think, like, obviously, some of them can take up a lot of context, but I think a lot of the MCP team has put a lot of work into this to, you know, make it better with tool search and things like that. So but, yeah, definitely call down MDs, like, any sort of instructions.
35:56The models just need, like, more room to run oftentimes.
35:59Yeah. Then do you, like instead of saying, like, know, you like, for example, let let's say writing a Twitter post. Like, instead of saying, like, hey.
36:05Make sure it's, like, two eighty characters or less, and do not do this. Like, do you give it, like, more, like, principles to follow?
36:12Like, you know, like, how how do you prompt it then? Like, is it more like
36:15Yeah. I I think the Twitter post is a good example, like, where you might give your give it context on yourself. So you're like, hey.
36:21I am you know, I work on Cloud Code. I'm Anthropic. This is sort of, like, some of, yeah, some of the principles we follow.
36:28Obviously, 280 characters is still, like, a good, like, sort of constraint because that's like, it does need to happen. Right?
36:35But you can even just say it's a tweet, and it will, like, sort of know like or this is a good example where you might say, like, hey. Keep it under 280 characters.
36:45But let's say that maybe there's a version of it that's better as a tweet thread with two tweets. You know what I mean? And so, like, in giving it the sense of, like, hey.
36:53I want to write a tweet thread. I'd prefer it to be one tweet. You know?
36:57Gives it more of optimism and flexibility, yeah, to to find something good. Okay.
37:02Got it. Okay.
37:03And and then let me ask this. Yeah. So Boris has been saying, like, coding is, a solved problem.
37:10Right? And and, like Yeah.
37:12Yeah. And and I feel like there's, like, becoming more technical these days is, like, a little bit like, I I guess my question is for someone like me who wants to actually, you know, learn how to work with agents more to actually learn what they're doing, Like, how do you actually become more technical? Like like, it's not about the syntax or anything.
37:27Right? It says is it just, like, using it more, or, like, how do you Yeah. I mean, I think for the first step is motivating yourself to learn things.
37:34You know what mean? I I think that is, like,
37:37um, honestly, really hard.
37:40And and, like, I find myself doing this as well where, like, you know, if you don't need to learn something to get the job done, maybe you won't. Right?
37:48But I think that, like, um, it is really important. I think the goal of learning to be more technical is to know my unknown unknowns.
37:57Right? And so there are some things that like, being technical and knowing the syntax of TypeScript is not really very helpful. You know?
38:05But I think, like, being technical and knowing, like, okay. Like, hey. What are the, like, trade offs of different back end services, and and what are the different video encryption libraries, and how do they work?
38:16And what's the difference between a local or remote video encryption or transcription library? Right?
38:21These are all, like, I think, pretty helpful. Right? And so I think oftentimes, I'm trying to learn the constraints of the system.
38:28Right? And so, like, what is possible? How is it doing it right now?
38:32How good could it be? You know? Like, what if we did something else?
38:35And, um, Claude can often brainstorm and teach you this if you, like, push it. Right? But you really do have to push it.
38:42I think, like, this is, like, you know, just the hardest thing about education and what everyone says is sort of, it feels good to, like, try and learn something, but, like, actually learning something is more work. You know?
38:54And, like, it should feel like work. I think Kaparti says this a lot. Right?
38:58Like, education should feel like work more than fun. So yeah. Well, we're all yeah.
39:02That's that's actually really tricky.
39:04Because because it's actually easier to just like, for example, the videos thing. Right?
39:07Like, it's actually easier to just keep prompting Cloud to just do it and then look at the output and then, like like, not actually learn anything, dude. Like, just, like, it's, go figure it out, and then it figures out. But the fact that you actually generate these, like, pretty detailed HTML reports and read them, I think it's probably the exception to that.
39:22I I think most people don't do this. Yeah.
39:25I mean, I think it's something I'm trying to push. Yeah. But, like, I think that, like, this is how you can learn to make just like I think the question for you is, like, okay.
39:33Like, how do you go from making good shorts to making, like, really, like, the best shorts, like, highest production quality shorts out there? You know what I mean? And I think if that's the goal, right, like, I think you need to, like, then sort of probably learn more about, like, both about video production and editing and, uh, and, like, the technical concepts, right, and then then push yourself that way.
39:55And I think that, like, I think we all want to push ourselves towards, like, being better and faster and not just faster. Yeah. So Okay.
40:02So maybe I'll add some custom instructions to my cloud.
40:05Be like, hey. Make sure you generate HTML reports for everything so I can read them. Yeah.
40:09Yeah. Yeah. Well, I mean, you know, yeah, I think it's
40:12something you want to push yourself to as well, you know, to, like, figure out when are you when do you want to understand if something is working. Got it. Awesome, man.
40:19Thank you for giving the gift of Cloud Code to all of us. And,
40:23yeah, really excited for what what's next, dude? I don't know if Yeah. Probably can share anything, but, you know, really excited.
40:27What's next? A lot of what's next is Cloud Tag. You know?
40:30And it'll keep getting better, but I I I think it's just something that it's hard for us to understate, like, how much it's sort of changed to work at Anthropic here. And so excited for everyone else to to get on board or to try it out.
40:42Yeah. If you have a really capable employee, you're not gonna, like, micromanage them. Right?
40:46You just say, hey. Take them on Slack, and they can't do this. Then, hopefully, it's just fun.
40:50Or you want to, like, talk talk with a iteratively with other people and collaborate. Yeah. You should build it to a point where I can walk by Claude's desk and just, like, ask her some questions and respond.
41:00Yeah. Like a robot. Yeah.
41:01Yeah. I mean, that could be a hack project for you. You know?
41:03I feel like I could already do that. Yeah.
41:05Cool. Alright, Farik. Well, um, really great chatting.
41:08And, um, yeah, I I I I think people will know where to find you online, so I don't think we have to talk about that. Yeah. Sounds good.
41:13Sounds good. Yeah. Amazing.
41:15Thanks, Peter. Appreciate it.
The Hook
The bait, then the rug-pull.
Thariq Shihipar works on the Claude Code team at Anthropic, and this conversation is a rare look at how the people building the tool actually use it: locking in exit conditions with /goal, splitting worker and verifier sub-agents to fight a model's bias toward grading its own work kindly, and treating a spec less like a document and more like a living thing that gets rewritten as the agent uncovers what nobody planned for.
Named ideas worth stealing.
01:30concept
/goal — locking the exit condition
/goal reminds the agent what its exit condition is and blocks it from stopping early. It's meant for tasks the user has already scoped and explored — the agent is told to push through and fill gaps itself rather than pausing to ask.
Steal forAny long build/refactor you've already scoped where you want zero check-ins until it's actually done.
28:20model
Workflows — worker + verifier sub-agents
- main agent decides the sub-tasks (e.g. which 5 clips to cut)
- each sub-agent does the work in parallel
- a separate verifier sub-agent grades the output against a rubric
- package the workflow as a JS file inside a reusable skill
Workflows spin up sub-agents to both parallelize work and verify it, splitting the roles specifically to counteract self-preferential bias — a model grading its own output is too lenient.
Steal forAny output with a non-deterministic 'is this good?' check — short-form clips, marketing copy variants, design directions.
08:40concept
Planning as unknown-removal
Planning isn't a one-time document handoff — it's an iterative loop of exploring, learning how underlying systems work (e.g. Whisper's failure modes), building a cheap prototype, and letting implementation surface what wasn't anticipated before respec'ing.
Steal forAny project where the spec keeps changing because you didn't know what you didn't know at the start.
34:10concept
System prompt cut 80% — fewer examples, fewer 'nevers'
As models got smarter, Anthropic found long example lists and blanket 'never' constraints actually over-constrain behavior. Trimming the prompt and explaining the reason behind a constraint (instead of a hard rule) gave the model more room to do the right thing.
Steal forAny CLAUDE.md, skill, or system prompt that has grown long — audit it for examples and 'never' rules that could become a stated reason instead.
How they asked for the click.
VERBAL ASK
14:15product
“Try it free at whisperflow.com, and use my code peter whisperflow to get six months free.”
Mid-roll host-read ad for WhisperFlow, delivered in Peter's own voice mid-conversation rather than as a separate produced spot.
MENTIONED ON CAMERA
Visual structure at a glance.
